5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 extern const struct regexp_engine my_reg_engine;
89 #include "dquote_static.c"
90 #include "charclass_invlists.h"
91 #include "inline_invlist.c"
92 #include "unicode_constants.h"
94 #define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
96 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
103 # if defined(BUGGY_MSC6)
104 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
105 # pragma optimize("a",off)
106 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
107 # pragma optimize("w",on )
108 # endif /* BUGGY_MSC6 */
112 #define STATIC static
116 typedef struct RExC_state_t {
117 U32 flags; /* RXf_* are we folding, multilining? */
118 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
119 char *precomp; /* uncompiled string. */
120 REGEXP *rx_sv; /* The SV that is the regexp. */
121 regexp *rx; /* perl core regexp structure */
122 regexp_internal *rxi; /* internal data for regexp object pprivate field */
123 char *start; /* Start of input for compile */
124 char *end; /* End of input for compile */
125 char *parse; /* Input-scan pointer. */
126 I32 whilem_seen; /* number of WHILEM in this expr */
127 regnode *emit_start; /* Start of emitted-code area */
128 regnode *emit_bound; /* First regnode outside of the allocated space */
129 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
130 implies compiling, so don't emit */
131 regnode emit_dummy; /* placeholder for emit to point to */
132 I32 naughty; /* How bad is this pattern? */
133 I32 sawback; /* Did we see \1, ...? */
135 I32 size; /* Code size. */
136 I32 npar; /* Capture buffer count, (OPEN). */
137 I32 cpar; /* Capture buffer count, (CLOSE). */
138 I32 nestroot; /* root parens we are in - used by accept */
141 regnode **open_parens; /* pointers to open parens */
142 regnode **close_parens; /* pointers to close parens */
143 regnode *opend; /* END node in program */
144 I32 utf8; /* whether the pattern is utf8 or not */
145 I32 orig_utf8; /* whether the pattern was originally in utf8 */
146 /* XXX use this for future optimisation of case
147 * where pattern must be upgraded to utf8. */
148 I32 uni_semantics; /* If a d charset modifier should use unicode
149 rules, even if the pattern is not in
151 HV *paren_names; /* Paren names */
153 regnode **recurse; /* Recurse regops */
154 I32 recurse_count; /* Number of recurse regops */
157 I32 override_recoding;
158 I32 in_multi_char_class;
159 struct reg_code_block *code_blocks; /* positions of literal (?{})
161 int num_code_blocks; /* size of code_blocks[] */
162 int code_index; /* next code_blocks[] slot */
164 char *starttry; /* -Dr: where regtry was called. */
165 #define RExC_starttry (pRExC_state->starttry)
167 SV *runtime_code_qr; /* qr with the runtime code blocks */
169 const char *lastparse;
171 AV *paren_name_list; /* idx -> name */
172 #define RExC_lastparse (pRExC_state->lastparse)
173 #define RExC_lastnum (pRExC_state->lastnum)
174 #define RExC_paren_name_list (pRExC_state->paren_name_list)
178 #define RExC_flags (pRExC_state->flags)
179 #define RExC_pm_flags (pRExC_state->pm_flags)
180 #define RExC_precomp (pRExC_state->precomp)
181 #define RExC_rx_sv (pRExC_state->rx_sv)
182 #define RExC_rx (pRExC_state->rx)
183 #define RExC_rxi (pRExC_state->rxi)
184 #define RExC_start (pRExC_state->start)
185 #define RExC_end (pRExC_state->end)
186 #define RExC_parse (pRExC_state->parse)
187 #define RExC_whilem_seen (pRExC_state->whilem_seen)
188 #ifdef RE_TRACK_PATTERN_OFFSETS
189 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
191 #define RExC_emit (pRExC_state->emit)
192 #define RExC_emit_dummy (pRExC_state->emit_dummy)
193 #define RExC_emit_start (pRExC_state->emit_start)
194 #define RExC_emit_bound (pRExC_state->emit_bound)
195 #define RExC_naughty (pRExC_state->naughty)
196 #define RExC_sawback (pRExC_state->sawback)
197 #define RExC_seen (pRExC_state->seen)
198 #define RExC_size (pRExC_state->size)
199 #define RExC_npar (pRExC_state->npar)
200 #define RExC_nestroot (pRExC_state->nestroot)
201 #define RExC_extralen (pRExC_state->extralen)
202 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
203 #define RExC_utf8 (pRExC_state->utf8)
204 #define RExC_uni_semantics (pRExC_state->uni_semantics)
205 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
206 #define RExC_open_parens (pRExC_state->open_parens)
207 #define RExC_close_parens (pRExC_state->close_parens)
208 #define RExC_opend (pRExC_state->opend)
209 #define RExC_paren_names (pRExC_state->paren_names)
210 #define RExC_recurse (pRExC_state->recurse)
211 #define RExC_recurse_count (pRExC_state->recurse_count)
212 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
213 #define RExC_contains_locale (pRExC_state->contains_locale)
214 #define RExC_override_recoding (pRExC_state->override_recoding)
215 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
218 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
219 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
220 ((*s) == '{' && regcurly(s, FALSE)))
223 #undef SPSTART /* dratted cpp namespace... */
226 * Flags to be passed up and down.
228 #define WORST 0 /* Worst case. */
229 #define HASWIDTH 0x01 /* Known to match non-null strings. */
231 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
232 * character. (There needs to be a case: in the switch statement in regexec.c
233 * for any node marked SIMPLE.) Note that this is not the same thing as
236 #define SPSTART 0x04 /* Starts with * or + */
237 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
238 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
239 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
241 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
243 /* whether trie related optimizations are enabled */
244 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
245 #define TRIE_STUDY_OPT
246 #define FULL_TRIE_STUDY
252 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
253 #define PBITVAL(paren) (1 << ((paren) & 7))
254 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
255 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
256 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
258 #define REQUIRE_UTF8 STMT_START { \
260 *flagp = RESTART_UTF8; \
265 /* This converts the named class defined in regcomp.h to its equivalent class
266 * number defined in handy.h. */
267 #define namedclass_to_classnum(class) ((int) ((class) / 2))
268 #define classnum_to_namedclass(classnum) ((classnum) * 2)
270 /* About scan_data_t.
272 During optimisation we recurse through the regexp program performing
273 various inplace (keyhole style) optimisations. In addition study_chunk
274 and scan_commit populate this data structure with information about
275 what strings MUST appear in the pattern. We look for the longest
276 string that must appear at a fixed location, and we look for the
277 longest string that may appear at a floating location. So for instance
282 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
283 strings (because they follow a .* construct). study_chunk will identify
284 both FOO and BAR as being the longest fixed and floating strings respectively.
286 The strings can be composites, for instance
290 will result in a composite fixed substring 'foo'.
292 For each string some basic information is maintained:
294 - offset or min_offset
295 This is the position the string must appear at, or not before.
296 It also implicitly (when combined with minlenp) tells us how many
297 characters must match before the string we are searching for.
298 Likewise when combined with minlenp and the length of the string it
299 tells us how many characters must appear after the string we have
303 Only used for floating strings. This is the rightmost point that
304 the string can appear at. If set to I32 max it indicates that the
305 string can occur infinitely far to the right.
308 A pointer to the minimum number of characters of the pattern that the
309 string was found inside. This is important as in the case of positive
310 lookahead or positive lookbehind we can have multiple patterns
315 The minimum length of the pattern overall is 3, the minimum length
316 of the lookahead part is 3, but the minimum length of the part that
317 will actually match is 1. So 'FOO's minimum length is 3, but the
318 minimum length for the F is 1. This is important as the minimum length
319 is used to determine offsets in front of and behind the string being
320 looked for. Since strings can be composites this is the length of the
321 pattern at the time it was committed with a scan_commit. Note that
322 the length is calculated by study_chunk, so that the minimum lengths
323 are not known until the full pattern has been compiled, thus the
324 pointer to the value.
328 In the case of lookbehind the string being searched for can be
329 offset past the start point of the final matching string.
330 If this value was just blithely removed from the min_offset it would
331 invalidate some of the calculations for how many chars must match
332 before or after (as they are derived from min_offset and minlen and
333 the length of the string being searched for).
334 When the final pattern is compiled and the data is moved from the
335 scan_data_t structure into the regexp structure the information
336 about lookbehind is factored in, with the information that would
337 have been lost precalculated in the end_shift field for the
340 The fields pos_min and pos_delta are used to store the minimum offset
341 and the delta to the maximum offset at the current point in the pattern.
345 typedef struct scan_data_t {
346 /*I32 len_min; unused */
347 /*I32 len_delta; unused */
351 I32 last_end; /* min value, <0 unless valid. */
354 SV **longest; /* Either &l_fixed, or &l_float. */
355 SV *longest_fixed; /* longest fixed string found in pattern */
356 I32 offset_fixed; /* offset where it starts */
357 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
358 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
359 SV *longest_float; /* longest floating string found in pattern */
360 I32 offset_float_min; /* earliest point in string it can appear */
361 I32 offset_float_max; /* latest point in string it can appear */
362 I32 *minlen_float; /* pointer to the minlen relevant to the string */
363 I32 lookbehind_float; /* is the position of the string modified by LB */
367 struct regnode_charclass_class *start_class;
371 * Forward declarations for pregcomp()'s friends.
374 static const scan_data_t zero_scan_data =
375 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
377 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
378 #define SF_BEFORE_SEOL 0x0001
379 #define SF_BEFORE_MEOL 0x0002
380 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
381 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
384 # define SF_FIX_SHIFT_EOL (0+2)
385 # define SF_FL_SHIFT_EOL (0+4)
387 # define SF_FIX_SHIFT_EOL (+2)
388 # define SF_FL_SHIFT_EOL (+4)
391 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
392 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
394 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
395 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
396 #define SF_IS_INF 0x0040
397 #define SF_HAS_PAR 0x0080
398 #define SF_IN_PAR 0x0100
399 #define SF_HAS_EVAL 0x0200
400 #define SCF_DO_SUBSTR 0x0400
401 #define SCF_DO_STCLASS_AND 0x0800
402 #define SCF_DO_STCLASS_OR 0x1000
403 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
404 #define SCF_WHILEM_VISITED_POS 0x2000
406 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
407 #define SCF_SEEN_ACCEPT 0x8000
408 #define SCF_TRIE_DOING_RESTUDY 0x10000
410 #define UTF cBOOL(RExC_utf8)
412 /* The enums for all these are ordered so things work out correctly */
413 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
414 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
415 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
416 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
417 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
418 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
419 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
421 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
423 #define OOB_NAMEDCLASS -1
425 /* There is no code point that is out-of-bounds, so this is problematic. But
426 * its only current use is to initialize a variable that is always set before
428 #define OOB_UNICODE 0xDEADBEEF
430 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
431 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
434 /* length of regex to show in messages that don't mark a position within */
435 #define RegexLengthToShowInErrorMessages 127
438 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
439 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
440 * op/pragma/warn/regcomp.
442 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
443 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
445 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
448 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
449 * arg. Show regex, up to a maximum length. If it's too long, chop and add
452 #define _FAIL(code) STMT_START { \
453 const char *ellipses = ""; \
454 IV len = RExC_end - RExC_precomp; \
457 SAVEFREESV(RExC_rx_sv); \
458 if (len > RegexLengthToShowInErrorMessages) { \
459 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
460 len = RegexLengthToShowInErrorMessages - 10; \
466 #define FAIL(msg) _FAIL( \
467 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
468 msg, (int)len, RExC_precomp, ellipses))
470 #define FAIL2(msg,arg) _FAIL( \
471 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
472 arg, (int)len, RExC_precomp, ellipses))
475 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
477 #define Simple_vFAIL(m) STMT_START { \
478 const IV offset = RExC_parse - RExC_precomp; \
479 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
480 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
484 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
486 #define vFAIL(m) STMT_START { \
488 SAVEFREESV(RExC_rx_sv); \
493 * Like Simple_vFAIL(), but accepts two arguments.
495 #define Simple_vFAIL2(m,a1) STMT_START { \
496 const IV offset = RExC_parse - RExC_precomp; \
497 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
498 (int)offset, RExC_precomp, RExC_precomp + offset); \
502 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
504 #define vFAIL2(m,a1) STMT_START { \
506 SAVEFREESV(RExC_rx_sv); \
507 Simple_vFAIL2(m, a1); \
512 * Like Simple_vFAIL(), but accepts three arguments.
514 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
515 const IV offset = RExC_parse - RExC_precomp; \
516 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
517 (int)offset, RExC_precomp, RExC_precomp + offset); \
521 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
523 #define vFAIL3(m,a1,a2) STMT_START { \
525 SAVEFREESV(RExC_rx_sv); \
526 Simple_vFAIL3(m, a1, a2); \
530 * Like Simple_vFAIL(), but accepts four arguments.
532 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
533 const IV offset = RExC_parse - RExC_precomp; \
534 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
535 (int)offset, RExC_precomp, RExC_precomp + offset); \
538 #define vFAIL4(m,a1,a2,a3) STMT_START { \
540 SAVEFREESV(RExC_rx_sv); \
541 Simple_vFAIL4(m, a1, a2, a3); \
544 /* m is not necessarily a "literal string", in this macro */
545 #define reg_warn_non_literal_string(loc, m) STMT_START { \
546 const IV offset = loc - RExC_precomp; \
547 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
548 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
551 #define ckWARNreg(loc,m) STMT_START { \
552 const IV offset = loc - RExC_precomp; \
553 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
554 (int)offset, RExC_precomp, RExC_precomp + offset); \
557 #define vWARN_dep(loc, m) STMT_START { \
558 const IV offset = loc - RExC_precomp; \
559 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
560 (int)offset, RExC_precomp, RExC_precomp + offset); \
563 #define ckWARNdep(loc,m) STMT_START { \
564 const IV offset = loc - RExC_precomp; \
565 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
567 (int)offset, RExC_precomp, RExC_precomp + offset); \
570 #define ckWARNregdep(loc,m) STMT_START { \
571 const IV offset = loc - RExC_precomp; \
572 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
574 (int)offset, RExC_precomp, RExC_precomp + offset); \
577 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
578 const IV offset = loc - RExC_precomp; \
579 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
581 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
584 #define ckWARN2reg(loc, m, a1) STMT_START { \
585 const IV offset = loc - RExC_precomp; \
586 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
587 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
590 #define vWARN3(loc, m, a1, a2) STMT_START { \
591 const IV offset = loc - RExC_precomp; \
592 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
593 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
596 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
597 const IV offset = loc - RExC_precomp; \
598 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
599 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
602 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
603 const IV offset = loc - RExC_precomp; \
604 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
605 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
608 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
609 const IV offset = loc - RExC_precomp; \
610 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
611 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
614 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
615 const IV offset = loc - RExC_precomp; \
616 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
617 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
621 /* Allow for side effects in s */
622 #define REGC(c,s) STMT_START { \
623 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
626 /* Macros for recording node offsets. 20001227 mjd@plover.com
627 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
628 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
629 * Element 0 holds the number n.
630 * Position is 1 indexed.
632 #ifndef RE_TRACK_PATTERN_OFFSETS
633 #define Set_Node_Offset_To_R(node,byte)
634 #define Set_Node_Offset(node,byte)
635 #define Set_Cur_Node_Offset
636 #define Set_Node_Length_To_R(node,len)
637 #define Set_Node_Length(node,len)
638 #define Set_Node_Cur_Length(node,start)
639 #define Node_Offset(n)
640 #define Node_Length(n)
641 #define Set_Node_Offset_Length(node,offset,len)
642 #define ProgLen(ri) ri->u.proglen
643 #define SetProgLen(ri,x) ri->u.proglen = x
645 #define ProgLen(ri) ri->u.offsets[0]
646 #define SetProgLen(ri,x) ri->u.offsets[0] = x
647 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
649 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
650 __LINE__, (int)(node), (int)(byte))); \
652 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
654 RExC_offsets[2*(node)-1] = (byte); \
659 #define Set_Node_Offset(node,byte) \
660 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
661 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
663 #define Set_Node_Length_To_R(node,len) STMT_START { \
665 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
666 __LINE__, (int)(node), (int)(len))); \
668 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
670 RExC_offsets[2*(node)] = (len); \
675 #define Set_Node_Length(node,len) \
676 Set_Node_Length_To_R((node)-RExC_emit_start, len)
677 #define Set_Node_Cur_Length(node, start) \
678 Set_Node_Length(node, RExC_parse - start)
680 /* Get offsets and lengths */
681 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
682 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
684 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
685 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
686 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
690 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
691 #define EXPERIMENTAL_INPLACESCAN
692 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
694 #define DEBUG_STUDYDATA(str,data,depth) \
695 DEBUG_OPTIMISE_MORE_r(if(data){ \
696 PerlIO_printf(Perl_debug_log, \
697 "%*s" str "Pos:%"IVdf"/%"IVdf \
698 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
699 (int)(depth)*2, "", \
700 (IV)((data)->pos_min), \
701 (IV)((data)->pos_delta), \
702 (UV)((data)->flags), \
703 (IV)((data)->whilem_c), \
704 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
705 is_inf ? "INF " : "" \
707 if ((data)->last_found) \
708 PerlIO_printf(Perl_debug_log, \
709 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
710 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
711 SvPVX_const((data)->last_found), \
712 (IV)((data)->last_end), \
713 (IV)((data)->last_start_min), \
714 (IV)((data)->last_start_max), \
715 ((data)->longest && \
716 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
717 SvPVX_const((data)->longest_fixed), \
718 (IV)((data)->offset_fixed), \
719 ((data)->longest && \
720 (data)->longest==&((data)->longest_float)) ? "*" : "", \
721 SvPVX_const((data)->longest_float), \
722 (IV)((data)->offset_float_min), \
723 (IV)((data)->offset_float_max) \
725 PerlIO_printf(Perl_debug_log,"\n"); \
728 /* Mark that we cannot extend a found fixed substring at this point.
729 Update the longest found anchored substring and the longest found
730 floating substrings if needed. */
733 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
735 const STRLEN l = CHR_SVLEN(data->last_found);
736 const STRLEN old_l = CHR_SVLEN(*data->longest);
737 GET_RE_DEBUG_FLAGS_DECL;
739 PERL_ARGS_ASSERT_SCAN_COMMIT;
741 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
742 SvSetMagicSV(*data->longest, data->last_found);
743 if (*data->longest == data->longest_fixed) {
744 data->offset_fixed = l ? data->last_start_min : data->pos_min;
745 if (data->flags & SF_BEFORE_EOL)
747 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
749 data->flags &= ~SF_FIX_BEFORE_EOL;
750 data->minlen_fixed=minlenp;
751 data->lookbehind_fixed=0;
753 else { /* *data->longest == data->longest_float */
754 data->offset_float_min = l ? data->last_start_min : data->pos_min;
755 data->offset_float_max = (l
756 ? data->last_start_max
757 : (data->pos_delta == I32_MAX ? I32_MAX : data->pos_min + data->pos_delta));
758 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
759 data->offset_float_max = I32_MAX;
760 if (data->flags & SF_BEFORE_EOL)
762 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
764 data->flags &= ~SF_FL_BEFORE_EOL;
765 data->minlen_float=minlenp;
766 data->lookbehind_float=0;
769 SvCUR_set(data->last_found, 0);
771 SV * const sv = data->last_found;
772 if (SvUTF8(sv) && SvMAGICAL(sv)) {
773 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
779 data->flags &= ~SF_BEFORE_EOL;
780 DEBUG_STUDYDATA("commit: ",data,0);
783 /* These macros set, clear and test whether the synthetic start class ('ssc',
784 * given by the parameter) matches an empty string (EOS). This uses the
785 * 'next_off' field in the node, to save a bit in the flags field. The ssc
786 * stands alone, so there is never a next_off, so this field is otherwise
787 * unused. The EOS information is used only for compilation, but theoretically
788 * it could be passed on to the execution code. This could be used to store
789 * more than one bit of information, but only this one is currently used. */
790 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
791 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
792 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
794 /* Can match anything (initialization) */
796 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
798 PERL_ARGS_ASSERT_CL_ANYTHING;
800 ANYOF_BITMAP_SETALL(cl);
801 cl->flags = ANYOF_UNICODE_ALL;
804 /* If any portion of the regex is to operate under locale rules,
805 * initialization includes it. The reason this isn't done for all regexes
806 * is that the optimizer was written under the assumption that locale was
807 * all-or-nothing. Given the complexity and lack of documentation in the
808 * optimizer, and that there are inadequate test cases for locale, so many
809 * parts of it may not work properly, it is safest to avoid locale unless
811 if (RExC_contains_locale) {
812 ANYOF_CLASS_SETALL(cl); /* /l uses class */
813 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
816 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
820 /* Can match anything (initialization) */
822 S_cl_is_anything(const struct regnode_charclass_class *cl)
826 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
828 for (value = 0; value < ANYOF_MAX; value += 2)
829 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
831 if (!(cl->flags & ANYOF_UNICODE_ALL))
833 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
838 /* Can match anything (initialization) */
840 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
842 PERL_ARGS_ASSERT_CL_INIT;
844 Zero(cl, 1, struct regnode_charclass_class);
846 cl_anything(pRExC_state, cl);
847 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
850 /* These two functions currently do the exact same thing */
851 #define cl_init_zero cl_init
853 /* 'AND' a given class with another one. Can create false positives. 'cl'
854 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
855 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
857 S_cl_and(struct regnode_charclass_class *cl,
858 const struct regnode_charclass_class *and_with)
860 PERL_ARGS_ASSERT_CL_AND;
862 assert(PL_regkind[and_with->type] == ANYOF);
864 /* I (khw) am not sure all these restrictions are necessary XXX */
865 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
866 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
867 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
868 && !(and_with->flags & ANYOF_LOC_FOLD)
869 && !(cl->flags & ANYOF_LOC_FOLD)) {
872 if (and_with->flags & ANYOF_INVERT)
873 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
874 cl->bitmap[i] &= ~and_with->bitmap[i];
876 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
877 cl->bitmap[i] &= and_with->bitmap[i];
878 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
880 if (and_with->flags & ANYOF_INVERT) {
882 /* Here, the and'ed node is inverted. Get the AND of the flags that
883 * aren't affected by the inversion. Those that are affected are
884 * handled individually below */
885 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
886 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
887 cl->flags |= affected_flags;
889 /* We currently don't know how to deal with things that aren't in the
890 * bitmap, but we know that the intersection is no greater than what
891 * is already in cl, so let there be false positives that get sorted
892 * out after the synthetic start class succeeds, and the node is
893 * matched for real. */
895 /* The inversion of these two flags indicate that the resulting
896 * intersection doesn't have them */
897 if (and_with->flags & ANYOF_UNICODE_ALL) {
898 cl->flags &= ~ANYOF_UNICODE_ALL;
900 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
901 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
904 else { /* and'd node is not inverted */
905 U8 outside_bitmap_but_not_utf8; /* Temp variable */
907 if (! ANYOF_NONBITMAP(and_with)) {
909 /* Here 'and_with' doesn't match anything outside the bitmap
910 * (except possibly ANYOF_UNICODE_ALL), which means the
911 * intersection can't either, except for ANYOF_UNICODE_ALL, in
912 * which case we don't know what the intersection is, but it's no
913 * greater than what cl already has, so can just leave it alone,
914 * with possible false positives */
915 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
916 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
917 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
920 else if (! ANYOF_NONBITMAP(cl)) {
922 /* Here, 'and_with' does match something outside the bitmap, and cl
923 * doesn't have a list of things to match outside the bitmap. If
924 * cl can match all code points above 255, the intersection will
925 * be those above-255 code points that 'and_with' matches. If cl
926 * can't match all Unicode code points, it means that it can't
927 * match anything outside the bitmap (since the 'if' that got us
928 * into this block tested for that), so we leave the bitmap empty.
930 if (cl->flags & ANYOF_UNICODE_ALL) {
931 ARG_SET(cl, ARG(and_with));
933 /* and_with's ARG may match things that don't require UTF8.
934 * And now cl's will too, in spite of this being an 'and'. See
935 * the comments below about the kludge */
936 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
940 /* Here, both 'and_with' and cl match something outside the
941 * bitmap. Currently we do not do the intersection, so just match
942 * whatever cl had at the beginning. */
946 /* Take the intersection of the two sets of flags. However, the
947 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
948 * kludge around the fact that this flag is not treated like the others
949 * which are initialized in cl_anything(). The way the optimizer works
950 * is that the synthetic start class (SSC) is initialized to match
951 * anything, and then the first time a real node is encountered, its
952 * values are AND'd with the SSC's with the result being the values of
953 * the real node. However, there are paths through the optimizer where
954 * the AND never gets called, so those initialized bits are set
955 * inappropriately, which is not usually a big deal, as they just cause
956 * false positives in the SSC, which will just mean a probably
957 * imperceptible slow down in execution. However this bit has a
958 * higher false positive consequence in that it can cause utf8.pm,
959 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
960 * bigger slowdown and also causes significant extra memory to be used.
961 * In order to prevent this, the code now takes a different tack. The
962 * bit isn't set unless some part of the regular expression needs it,
963 * but once set it won't get cleared. This means that these extra
964 * modules won't get loaded unless there was some path through the
965 * pattern that would have required them anyway, and so any false
966 * positives that occur by not ANDing them out when they could be
967 * aren't as severe as they would be if we treated this bit like all
969 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
970 & ANYOF_NONBITMAP_NON_UTF8;
971 cl->flags &= and_with->flags;
972 cl->flags |= outside_bitmap_but_not_utf8;
976 /* 'OR' a given class with another one. Can create false positives. 'cl'
977 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
978 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
980 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
982 PERL_ARGS_ASSERT_CL_OR;
984 if (or_with->flags & ANYOF_INVERT) {
986 /* Here, the or'd node is to be inverted. This means we take the
987 * complement of everything not in the bitmap, but currently we don't
988 * know what that is, so give up and match anything */
989 if (ANYOF_NONBITMAP(or_with)) {
990 cl_anything(pRExC_state, cl);
993 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
994 * <= (B1 | !B2) | (CL1 | !CL2)
995 * which is wasteful if CL2 is small, but we ignore CL2:
996 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
997 * XXXX Can we handle case-fold? Unclear:
998 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
999 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
1001 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1002 && !(or_with->flags & ANYOF_LOC_FOLD)
1003 && !(cl->flags & ANYOF_LOC_FOLD) ) {
1006 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1007 cl->bitmap[i] |= ~or_with->bitmap[i];
1008 } /* XXXX: logic is complicated otherwise */
1010 cl_anything(pRExC_state, cl);
1013 /* And, we can just take the union of the flags that aren't affected
1014 * by the inversion */
1015 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1017 /* For the remaining flags:
1018 ANYOF_UNICODE_ALL and inverted means to not match anything above
1019 255, which means that the union with cl should just be
1020 what cl has in it, so can ignore this flag
1021 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1022 is 127-255 to match them, but then invert that, so the
1023 union with cl should just be what cl has in it, so can
1026 } else { /* 'or_with' is not inverted */
1027 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1028 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1029 && (!(or_with->flags & ANYOF_LOC_FOLD)
1030 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1033 /* OR char bitmap and class bitmap separately */
1034 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1035 cl->bitmap[i] |= or_with->bitmap[i];
1036 if (or_with->flags & ANYOF_CLASS) {
1037 ANYOF_CLASS_OR(or_with, cl);
1040 else { /* XXXX: logic is complicated, leave it along for a moment. */
1041 cl_anything(pRExC_state, cl);
1044 if (ANYOF_NONBITMAP(or_with)) {
1046 /* Use the added node's outside-the-bit-map match if there isn't a
1047 * conflict. If there is a conflict (both nodes match something
1048 * outside the bitmap, but what they match outside is not the same
1049 * pointer, and hence not easily compared until XXX we extend
1050 * inversion lists this far), give up and allow the start class to
1051 * match everything outside the bitmap. If that stuff is all above
1052 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1053 if (! ANYOF_NONBITMAP(cl)) {
1054 ARG_SET(cl, ARG(or_with));
1056 else if (ARG(cl) != ARG(or_with)) {
1058 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1059 cl_anything(pRExC_state, cl);
1062 cl->flags |= ANYOF_UNICODE_ALL;
1067 /* Take the union */
1068 cl->flags |= or_with->flags;
1072 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1073 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1074 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1075 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1080 dump_trie(trie,widecharmap,revcharmap)
1081 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1082 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1084 These routines dump out a trie in a somewhat readable format.
1085 The _interim_ variants are used for debugging the interim
1086 tables that are used to generate the final compressed
1087 representation which is what dump_trie expects.
1089 Part of the reason for their existence is to provide a form
1090 of documentation as to how the different representations function.
1095 Dumps the final compressed table form of the trie to Perl_debug_log.
1096 Used for debugging make_trie().
1100 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1101 AV *revcharmap, U32 depth)
1104 SV *sv=sv_newmortal();
1105 int colwidth= widecharmap ? 6 : 4;
1107 GET_RE_DEBUG_FLAGS_DECL;
1109 PERL_ARGS_ASSERT_DUMP_TRIE;
1111 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1112 (int)depth * 2 + 2,"",
1113 "Match","Base","Ofs" );
1115 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1116 SV ** const tmp = av_fetch( revcharmap, state, 0);
1118 PerlIO_printf( Perl_debug_log, "%*s",
1120 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1121 PL_colors[0], PL_colors[1],
1122 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1123 PERL_PV_ESCAPE_FIRSTCHAR
1128 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1129 (int)depth * 2 + 2,"");
1131 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1132 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1133 PerlIO_printf( Perl_debug_log, "\n");
1135 for( state = 1 ; state < trie->statecount ; state++ ) {
1136 const U32 base = trie->states[ state ].trans.base;
1138 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1140 if ( trie->states[ state ].wordnum ) {
1141 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1143 PerlIO_printf( Perl_debug_log, "%6s", "" );
1146 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1151 while( ( base + ofs < trie->uniquecharcount ) ||
1152 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1153 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1156 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1158 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1159 if ( ( base + ofs >= trie->uniquecharcount ) &&
1160 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1161 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1163 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1165 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1167 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1171 PerlIO_printf( Perl_debug_log, "]");
1174 PerlIO_printf( Perl_debug_log, "\n" );
1176 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1177 for (word=1; word <= trie->wordcount; word++) {
1178 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1179 (int)word, (int)(trie->wordinfo[word].prev),
1180 (int)(trie->wordinfo[word].len));
1182 PerlIO_printf(Perl_debug_log, "\n" );
1185 Dumps a fully constructed but uncompressed trie in list form.
1186 List tries normally only are used for construction when the number of
1187 possible chars (trie->uniquecharcount) is very high.
1188 Used for debugging make_trie().
1191 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1192 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1196 SV *sv=sv_newmortal();
1197 int colwidth= widecharmap ? 6 : 4;
1198 GET_RE_DEBUG_FLAGS_DECL;
1200 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1202 /* print out the table precompression. */
1203 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1204 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1205 "------:-----+-----------------\n" );
1207 for( state=1 ; state < next_alloc ; state ++ ) {
1210 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1211 (int)depth * 2 + 2,"", (UV)state );
1212 if ( ! trie->states[ state ].wordnum ) {
1213 PerlIO_printf( Perl_debug_log, "%5s| ","");
1215 PerlIO_printf( Perl_debug_log, "W%4x| ",
1216 trie->states[ state ].wordnum
1219 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1220 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1222 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1224 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1225 PL_colors[0], PL_colors[1],
1226 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1227 PERL_PV_ESCAPE_FIRSTCHAR
1229 TRIE_LIST_ITEM(state,charid).forid,
1230 (UV)TRIE_LIST_ITEM(state,charid).newstate
1233 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1234 (int)((depth * 2) + 14), "");
1237 PerlIO_printf( Perl_debug_log, "\n");
1242 Dumps a fully constructed but uncompressed trie in table form.
1243 This is the normal DFA style state transition table, with a few
1244 twists to facilitate compression later.
1245 Used for debugging make_trie().
1248 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1249 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1254 SV *sv=sv_newmortal();
1255 int colwidth= widecharmap ? 6 : 4;
1256 GET_RE_DEBUG_FLAGS_DECL;
1258 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1261 print out the table precompression so that we can do a visual check
1262 that they are identical.
1265 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1267 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1268 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1270 PerlIO_printf( Perl_debug_log, "%*s",
1272 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1273 PL_colors[0], PL_colors[1],
1274 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1275 PERL_PV_ESCAPE_FIRSTCHAR
1281 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1283 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1284 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1287 PerlIO_printf( Perl_debug_log, "\n" );
1289 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1291 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1292 (int)depth * 2 + 2,"",
1293 (UV)TRIE_NODENUM( state ) );
1295 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1296 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1298 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1300 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1302 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1303 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1305 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1306 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1314 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1315 startbranch: the first branch in the whole branch sequence
1316 first : start branch of sequence of branch-exact nodes.
1317 May be the same as startbranch
1318 last : Thing following the last branch.
1319 May be the same as tail.
1320 tail : item following the branch sequence
1321 count : words in the sequence
1322 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1323 depth : indent depth
1325 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1327 A trie is an N'ary tree where the branches are determined by digital
1328 decomposition of the key. IE, at the root node you look up the 1st character and
1329 follow that branch repeat until you find the end of the branches. Nodes can be
1330 marked as "accepting" meaning they represent a complete word. Eg:
1334 would convert into the following structure. Numbers represent states, letters
1335 following numbers represent valid transitions on the letter from that state, if
1336 the number is in square brackets it represents an accepting state, otherwise it
1337 will be in parenthesis.
1339 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1343 (1) +-i->(6)-+-s->[7]
1345 +-s->(3)-+-h->(4)-+-e->[5]
1347 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1349 This shows that when matching against the string 'hers' we will begin at state 1
1350 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1351 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1352 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1353 single traverse. We store a mapping from accepting to state to which word was
1354 matched, and then when we have multiple possibilities we try to complete the
1355 rest of the regex in the order in which they occured in the alternation.
1357 The only prior NFA like behaviour that would be changed by the TRIE support is
1358 the silent ignoring of duplicate alternations which are of the form:
1360 / (DUPE|DUPE) X? (?{ ... }) Y /x
1362 Thus EVAL blocks following a trie may be called a different number of times with
1363 and without the optimisation. With the optimisations dupes will be silently
1364 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1365 the following demonstrates:
1367 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1369 which prints out 'word' three times, but
1371 'words'=~/(word|word|word)(?{ print $1 })S/
1373 which doesnt print it out at all. This is due to other optimisations kicking in.
1375 Example of what happens on a structural level:
1377 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1379 1: CURLYM[1] {1,32767}(18)
1390 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1391 and should turn into:
1393 1: CURLYM[1] {1,32767}(18)
1395 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1403 Cases where tail != last would be like /(?foo|bar)baz/:
1413 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1414 and would end up looking like:
1417 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1424 d = uvuni_to_utf8_flags(d, uv, 0);
1426 is the recommended Unicode-aware way of saying
1431 #define TRIE_STORE_REVCHAR(val) \
1434 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1435 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1436 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1437 SvCUR_set(zlopp, kapow - flrbbbbb); \
1440 av_push(revcharmap, zlopp); \
1442 char ooooff = (char)val; \
1443 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1447 #define TRIE_READ_CHAR STMT_START { \
1450 /* if it is UTF then it is either already folded, or does not need folding */ \
1451 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1453 else if (folder == PL_fold_latin1) { \
1454 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1455 if ( foldlen > 0 ) { \
1456 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1462 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, FOLD_FLAGS_FULL); \
1463 skiplen = UNISKIP(uvc); \
1464 foldlen -= skiplen; \
1465 scan = foldbuf + skiplen; \
1468 /* raw data, will be folded later if needed */ \
1476 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1477 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1478 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1479 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1481 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1482 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1483 TRIE_LIST_CUR( state )++; \
1486 #define TRIE_LIST_NEW(state) STMT_START { \
1487 Newxz( trie->states[ state ].trans.list, \
1488 4, reg_trie_trans_le ); \
1489 TRIE_LIST_CUR( state ) = 1; \
1490 TRIE_LIST_LEN( state ) = 4; \
1493 #define TRIE_HANDLE_WORD(state) STMT_START { \
1494 U16 dupe= trie->states[ state ].wordnum; \
1495 regnode * const noper_next = regnext( noper ); \
1498 /* store the word for dumping */ \
1500 if (OP(noper) != NOTHING) \
1501 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1503 tmp = newSVpvn_utf8( "", 0, UTF ); \
1504 av_push( trie_words, tmp ); \
1508 trie->wordinfo[curword].prev = 0; \
1509 trie->wordinfo[curword].len = wordlen; \
1510 trie->wordinfo[curword].accept = state; \
1512 if ( noper_next < tail ) { \
1514 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1515 trie->jump[curword] = (U16)(noper_next - convert); \
1517 jumper = noper_next; \
1519 nextbranch= regnext(cur); \
1523 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1524 /* chain, so that when the bits of chain are later */\
1525 /* linked together, the dups appear in the chain */\
1526 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1527 trie->wordinfo[dupe].prev = curword; \
1529 /* we haven't inserted this word yet. */ \
1530 trie->states[ state ].wordnum = curword; \
1535 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1536 ( ( base + charid >= ucharcount \
1537 && base + charid < ubound \
1538 && state == trie->trans[ base - ucharcount + charid ].check \
1539 && trie->trans[ base - ucharcount + charid ].next ) \
1540 ? trie->trans[ base - ucharcount + charid ].next \
1541 : ( state==1 ? special : 0 ) \
1545 #define MADE_JUMP_TRIE 2
1546 #define MADE_EXACT_TRIE 4
1549 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1552 /* first pass, loop through and scan words */
1553 reg_trie_data *trie;
1554 HV *widecharmap = NULL;
1555 AV *revcharmap = newAV();
1557 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1562 regnode *jumper = NULL;
1563 regnode *nextbranch = NULL;
1564 regnode *convert = NULL;
1565 U32 *prev_states; /* temp array mapping each state to previous one */
1566 /* we just use folder as a flag in utf8 */
1567 const U8 * folder = NULL;
1570 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1571 AV *trie_words = NULL;
1572 /* along with revcharmap, this only used during construction but both are
1573 * useful during debugging so we store them in the struct when debugging.
1576 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1577 STRLEN trie_charcount=0;
1579 SV *re_trie_maxbuff;
1580 GET_RE_DEBUG_FLAGS_DECL;
1582 PERL_ARGS_ASSERT_MAKE_TRIE;
1584 PERL_UNUSED_ARG(depth);
1591 case EXACTFU_TRICKYFOLD:
1592 case EXACTFU: folder = PL_fold_latin1; break;
1593 case EXACTF: folder = PL_fold; break;
1594 case EXACTFL: folder = PL_fold_locale; break;
1595 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1598 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1600 trie->startstate = 1;
1601 trie->wordcount = word_count;
1602 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1603 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1605 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1606 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1607 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1610 trie_words = newAV();
1613 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1614 if (!SvIOK(re_trie_maxbuff)) {
1615 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1617 DEBUG_TRIE_COMPILE_r({
1618 PerlIO_printf( Perl_debug_log,
1619 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1620 (int)depth * 2 + 2, "",
1621 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1622 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1626 /* Find the node we are going to overwrite */
1627 if ( first == startbranch && OP( last ) != BRANCH ) {
1628 /* whole branch chain */
1631 /* branch sub-chain */
1632 convert = NEXTOPER( first );
1635 /* -- First loop and Setup --
1637 We first traverse the branches and scan each word to determine if it
1638 contains widechars, and how many unique chars there are, this is
1639 important as we have to build a table with at least as many columns as we
1642 We use an array of integers to represent the character codes 0..255
1643 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1644 native representation of the character value as the key and IV's for the
1647 *TODO* If we keep track of how many times each character is used we can
1648 remap the columns so that the table compression later on is more
1649 efficient in terms of memory by ensuring the most common value is in the
1650 middle and the least common are on the outside. IMO this would be better
1651 than a most to least common mapping as theres a decent chance the most
1652 common letter will share a node with the least common, meaning the node
1653 will not be compressible. With a middle is most common approach the worst
1654 case is when we have the least common nodes twice.
1658 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1659 regnode *noper = NEXTOPER( cur );
1660 const U8 *uc = (U8*)STRING( noper );
1661 const U8 *e = uc + STR_LEN( noper );
1663 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1665 const U8 *scan = (U8*)NULL;
1666 U32 wordlen = 0; /* required init */
1668 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1670 if (OP(noper) == NOTHING) {
1671 regnode *noper_next= regnext(noper);
1672 if (noper_next != tail && OP(noper_next) == flags) {
1674 uc= (U8*)STRING(noper);
1675 e= uc + STR_LEN(noper);
1676 trie->minlen= STR_LEN(noper);
1683 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1684 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1685 regardless of encoding */
1686 if (OP( noper ) == EXACTFU_SS) {
1687 /* false positives are ok, so just set this */
1688 TRIE_BITMAP_SET(trie,0xDF);
1691 for ( ; uc < e ; uc += len ) {
1692 TRIE_CHARCOUNT(trie)++;
1697 U8 folded= folder[ (U8) uvc ];
1698 if ( !trie->charmap[ folded ] ) {
1699 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1700 TRIE_STORE_REVCHAR( folded );
1703 if ( !trie->charmap[ uvc ] ) {
1704 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1705 TRIE_STORE_REVCHAR( uvc );
1708 /* store the codepoint in the bitmap, and its folded
1710 TRIE_BITMAP_SET(trie, uvc);
1712 /* store the folded codepoint */
1713 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1716 /* store first byte of utf8 representation of
1717 variant codepoints */
1718 if (! UNI_IS_INVARIANT(uvc)) {
1719 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1722 set_bit = 0; /* We've done our bit :-) */
1727 widecharmap = newHV();
1729 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1732 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1734 if ( !SvTRUE( *svpp ) ) {
1735 sv_setiv( *svpp, ++trie->uniquecharcount );
1736 TRIE_STORE_REVCHAR(uvc);
1740 if( cur == first ) {
1741 trie->minlen = chars;
1742 trie->maxlen = chars;
1743 } else if (chars < trie->minlen) {
1744 trie->minlen = chars;
1745 } else if (chars > trie->maxlen) {
1746 trie->maxlen = chars;
1748 if (OP( noper ) == EXACTFU_SS) {
1749 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1750 if (trie->minlen > 1)
1753 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1754 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1755 * - We assume that any such sequence might match a 2 byte string */
1756 if (trie->minlen > 2 )
1760 } /* end first pass */
1761 DEBUG_TRIE_COMPILE_r(
1762 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1763 (int)depth * 2 + 2,"",
1764 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1765 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1766 (int)trie->minlen, (int)trie->maxlen )
1770 We now know what we are dealing with in terms of unique chars and
1771 string sizes so we can calculate how much memory a naive
1772 representation using a flat table will take. If it's over a reasonable
1773 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1774 conservative but potentially much slower representation using an array
1777 At the end we convert both representations into the same compressed
1778 form that will be used in regexec.c for matching with. The latter
1779 is a form that cannot be used to construct with but has memory
1780 properties similar to the list form and access properties similar
1781 to the table form making it both suitable for fast searches and
1782 small enough that its feasable to store for the duration of a program.
1784 See the comment in the code where the compressed table is produced
1785 inplace from the flat tabe representation for an explanation of how
1786 the compression works.
1791 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1794 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1796 Second Pass -- Array Of Lists Representation
1798 Each state will be represented by a list of charid:state records
1799 (reg_trie_trans_le) the first such element holds the CUR and LEN
1800 points of the allocated array. (See defines above).
1802 We build the initial structure using the lists, and then convert
1803 it into the compressed table form which allows faster lookups
1804 (but cant be modified once converted).
1807 STRLEN transcount = 1;
1809 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1810 "%*sCompiling trie using list compiler\n",
1811 (int)depth * 2 + 2, ""));
1813 trie->states = (reg_trie_state *)
1814 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1815 sizeof(reg_trie_state) );
1819 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1821 regnode *noper = NEXTOPER( cur );
1822 U8 *uc = (U8*)STRING( noper );
1823 const U8 *e = uc + STR_LEN( noper );
1824 U32 state = 1; /* required init */
1825 U16 charid = 0; /* sanity init */
1826 U8 *scan = (U8*)NULL; /* sanity init */
1827 STRLEN foldlen = 0; /* required init */
1828 U32 wordlen = 0; /* required init */
1829 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1832 if (OP(noper) == NOTHING) {
1833 regnode *noper_next= regnext(noper);
1834 if (noper_next != tail && OP(noper_next) == flags) {
1836 uc= (U8*)STRING(noper);
1837 e= uc + STR_LEN(noper);
1841 if (OP(noper) != NOTHING) {
1842 for ( ; uc < e ; uc += len ) {
1847 charid = trie->charmap[ uvc ];
1849 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1853 charid=(U16)SvIV( *svpp );
1856 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1863 if ( !trie->states[ state ].trans.list ) {
1864 TRIE_LIST_NEW( state );
1866 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1867 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1868 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1873 newstate = next_alloc++;
1874 prev_states[newstate] = state;
1875 TRIE_LIST_PUSH( state, charid, newstate );
1880 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1884 TRIE_HANDLE_WORD(state);
1886 } /* end second pass */
1888 /* next alloc is the NEXT state to be allocated */
1889 trie->statecount = next_alloc;
1890 trie->states = (reg_trie_state *)
1891 PerlMemShared_realloc( trie->states,
1893 * sizeof(reg_trie_state) );
1895 /* and now dump it out before we compress it */
1896 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1897 revcharmap, next_alloc,
1901 trie->trans = (reg_trie_trans *)
1902 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1909 for( state=1 ; state < next_alloc ; state ++ ) {
1913 DEBUG_TRIE_COMPILE_MORE_r(
1914 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1918 if (trie->states[state].trans.list) {
1919 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1923 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1924 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1925 if ( forid < minid ) {
1927 } else if ( forid > maxid ) {
1931 if ( transcount < tp + maxid - minid + 1) {
1933 trie->trans = (reg_trie_trans *)
1934 PerlMemShared_realloc( trie->trans,
1936 * sizeof(reg_trie_trans) );
1937 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1939 base = trie->uniquecharcount + tp - minid;
1940 if ( maxid == minid ) {
1942 for ( ; zp < tp ; zp++ ) {
1943 if ( ! trie->trans[ zp ].next ) {
1944 base = trie->uniquecharcount + zp - minid;
1945 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1946 trie->trans[ zp ].check = state;
1952 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1953 trie->trans[ tp ].check = state;
1958 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1959 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1960 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1961 trie->trans[ tid ].check = state;
1963 tp += ( maxid - minid + 1 );
1965 Safefree(trie->states[ state ].trans.list);
1968 DEBUG_TRIE_COMPILE_MORE_r(
1969 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1972 trie->states[ state ].trans.base=base;
1974 trie->lasttrans = tp + 1;
1978 Second Pass -- Flat Table Representation.
1980 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1981 We know that we will need Charcount+1 trans at most to store the data
1982 (one row per char at worst case) So we preallocate both structures
1983 assuming worst case.
1985 We then construct the trie using only the .next slots of the entry
1988 We use the .check field of the first entry of the node temporarily to
1989 make compression both faster and easier by keeping track of how many non
1990 zero fields are in the node.
1992 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1995 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1996 number representing the first entry of the node, and state as a
1997 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1998 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1999 are 2 entrys per node. eg:
2007 The table is internally in the right hand, idx form. However as we also
2008 have to deal with the states array which is indexed by nodenum we have to
2009 use TRIE_NODENUM() to convert.
2012 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2013 "%*sCompiling trie using table compiler\n",
2014 (int)depth * 2 + 2, ""));
2016 trie->trans = (reg_trie_trans *)
2017 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2018 * trie->uniquecharcount + 1,
2019 sizeof(reg_trie_trans) );
2020 trie->states = (reg_trie_state *)
2021 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2022 sizeof(reg_trie_state) );
2023 next_alloc = trie->uniquecharcount + 1;
2026 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2028 regnode *noper = NEXTOPER( cur );
2029 const U8 *uc = (U8*)STRING( noper );
2030 const U8 *e = uc + STR_LEN( noper );
2032 U32 state = 1; /* required init */
2034 U16 charid = 0; /* sanity init */
2035 U32 accept_state = 0; /* sanity init */
2036 U8 *scan = (U8*)NULL; /* sanity init */
2038 STRLEN foldlen = 0; /* required init */
2039 U32 wordlen = 0; /* required init */
2041 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2043 if (OP(noper) == NOTHING) {
2044 regnode *noper_next= regnext(noper);
2045 if (noper_next != tail && OP(noper_next) == flags) {
2047 uc= (U8*)STRING(noper);
2048 e= uc + STR_LEN(noper);
2052 if ( OP(noper) != NOTHING ) {
2053 for ( ; uc < e ; uc += len ) {
2058 charid = trie->charmap[ uvc ];
2060 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2061 charid = svpp ? (U16)SvIV(*svpp) : 0;
2065 if ( !trie->trans[ state + charid ].next ) {
2066 trie->trans[ state + charid ].next = next_alloc;
2067 trie->trans[ state ].check++;
2068 prev_states[TRIE_NODENUM(next_alloc)]
2069 = TRIE_NODENUM(state);
2070 next_alloc += trie->uniquecharcount;
2072 state = trie->trans[ state + charid ].next;
2074 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2076 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2079 accept_state = TRIE_NODENUM( state );
2080 TRIE_HANDLE_WORD(accept_state);
2082 } /* end second pass */
2084 /* and now dump it out before we compress it */
2085 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2087 next_alloc, depth+1));
2091 * Inplace compress the table.*
2093 For sparse data sets the table constructed by the trie algorithm will
2094 be mostly 0/FAIL transitions or to put it another way mostly empty.
2095 (Note that leaf nodes will not contain any transitions.)
2097 This algorithm compresses the tables by eliminating most such
2098 transitions, at the cost of a modest bit of extra work during lookup:
2100 - Each states[] entry contains a .base field which indicates the
2101 index in the state[] array wheres its transition data is stored.
2103 - If .base is 0 there are no valid transitions from that node.
2105 - If .base is nonzero then charid is added to it to find an entry in
2108 -If trans[states[state].base+charid].check!=state then the
2109 transition is taken to be a 0/Fail transition. Thus if there are fail
2110 transitions at the front of the node then the .base offset will point
2111 somewhere inside the previous nodes data (or maybe even into a node
2112 even earlier), but the .check field determines if the transition is
2116 The following process inplace converts the table to the compressed
2117 table: We first do not compress the root node 1,and mark all its
2118 .check pointers as 1 and set its .base pointer as 1 as well. This
2119 allows us to do a DFA construction from the compressed table later,
2120 and ensures that any .base pointers we calculate later are greater
2123 - We set 'pos' to indicate the first entry of the second node.
2125 - We then iterate over the columns of the node, finding the first and
2126 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2127 and set the .check pointers accordingly, and advance pos
2128 appropriately and repreat for the next node. Note that when we copy
2129 the next pointers we have to convert them from the original
2130 NODEIDX form to NODENUM form as the former is not valid post
2133 - If a node has no transitions used we mark its base as 0 and do not
2134 advance the pos pointer.
2136 - If a node only has one transition we use a second pointer into the
2137 structure to fill in allocated fail transitions from other states.
2138 This pointer is independent of the main pointer and scans forward
2139 looking for null transitions that are allocated to a state. When it
2140 finds one it writes the single transition into the "hole". If the
2141 pointer doesnt find one the single transition is appended as normal.
2143 - Once compressed we can Renew/realloc the structures to release the
2146 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2147 specifically Fig 3.47 and the associated pseudocode.
2151 const U32 laststate = TRIE_NODENUM( next_alloc );
2154 trie->statecount = laststate;
2156 for ( state = 1 ; state < laststate ; state++ ) {
2158 const U32 stateidx = TRIE_NODEIDX( state );
2159 const U32 o_used = trie->trans[ stateidx ].check;
2160 U32 used = trie->trans[ stateidx ].check;
2161 trie->trans[ stateidx ].check = 0;
2163 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2164 if ( flag || trie->trans[ stateidx + charid ].next ) {
2165 if ( trie->trans[ stateidx + charid ].next ) {
2167 for ( ; zp < pos ; zp++ ) {
2168 if ( ! trie->trans[ zp ].next ) {
2172 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2173 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2174 trie->trans[ zp ].check = state;
2175 if ( ++zp > pos ) pos = zp;
2182 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2184 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2185 trie->trans[ pos ].check = state;
2190 trie->lasttrans = pos + 1;
2191 trie->states = (reg_trie_state *)
2192 PerlMemShared_realloc( trie->states, laststate
2193 * sizeof(reg_trie_state) );
2194 DEBUG_TRIE_COMPILE_MORE_r(
2195 PerlIO_printf( Perl_debug_log,
2196 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2197 (int)depth * 2 + 2,"",
2198 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2201 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2204 } /* end table compress */
2206 DEBUG_TRIE_COMPILE_MORE_r(
2207 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2208 (int)depth * 2 + 2, "",
2209 (UV)trie->statecount,
2210 (UV)trie->lasttrans)
2212 /* resize the trans array to remove unused space */
2213 trie->trans = (reg_trie_trans *)
2214 PerlMemShared_realloc( trie->trans, trie->lasttrans
2215 * sizeof(reg_trie_trans) );
2217 { /* Modify the program and insert the new TRIE node */
2218 U8 nodetype =(U8)(flags & 0xFF);
2222 regnode *optimize = NULL;
2223 #ifdef RE_TRACK_PATTERN_OFFSETS
2226 U32 mjd_nodelen = 0;
2227 #endif /* RE_TRACK_PATTERN_OFFSETS */
2228 #endif /* DEBUGGING */
2230 This means we convert either the first branch or the first Exact,
2231 depending on whether the thing following (in 'last') is a branch
2232 or not and whther first is the startbranch (ie is it a sub part of
2233 the alternation or is it the whole thing.)
2234 Assuming its a sub part we convert the EXACT otherwise we convert
2235 the whole branch sequence, including the first.
2237 /* Find the node we are going to overwrite */
2238 if ( first != startbranch || OP( last ) == BRANCH ) {
2239 /* branch sub-chain */
2240 NEXT_OFF( first ) = (U16)(last - first);
2241 #ifdef RE_TRACK_PATTERN_OFFSETS
2243 mjd_offset= Node_Offset((convert));
2244 mjd_nodelen= Node_Length((convert));
2247 /* whole branch chain */
2249 #ifdef RE_TRACK_PATTERN_OFFSETS
2252 const regnode *nop = NEXTOPER( convert );
2253 mjd_offset= Node_Offset((nop));
2254 mjd_nodelen= Node_Length((nop));
2258 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2259 (int)depth * 2 + 2, "",
2260 (UV)mjd_offset, (UV)mjd_nodelen)
2263 /* But first we check to see if there is a common prefix we can
2264 split out as an EXACT and put in front of the TRIE node. */
2265 trie->startstate= 1;
2266 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2268 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2272 const U32 base = trie->states[ state ].trans.base;
2274 if ( trie->states[state].wordnum )
2277 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2278 if ( ( base + ofs >= trie->uniquecharcount ) &&
2279 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2280 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2282 if ( ++count > 1 ) {
2283 SV **tmp = av_fetch( revcharmap, ofs, 0);
2284 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2285 if ( state == 1 ) break;
2287 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2289 PerlIO_printf(Perl_debug_log,
2290 "%*sNew Start State=%"UVuf" Class: [",
2291 (int)depth * 2 + 2, "",
2294 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2295 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2297 TRIE_BITMAP_SET(trie,*ch);
2299 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2301 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2305 TRIE_BITMAP_SET(trie,*ch);
2307 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2308 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2314 SV **tmp = av_fetch( revcharmap, idx, 0);
2316 char *ch = SvPV( *tmp, len );
2318 SV *sv=sv_newmortal();
2319 PerlIO_printf( Perl_debug_log,
2320 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2321 (int)depth * 2 + 2, "",
2323 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2324 PL_colors[0], PL_colors[1],
2325 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2326 PERL_PV_ESCAPE_FIRSTCHAR
2331 OP( convert ) = nodetype;
2332 str=STRING(convert);
2335 STR_LEN(convert) += len;
2341 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2346 trie->prefixlen = (state-1);
2348 regnode *n = convert+NODE_SZ_STR(convert);
2349 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2350 trie->startstate = state;
2351 trie->minlen -= (state - 1);
2352 trie->maxlen -= (state - 1);
2354 /* At least the UNICOS C compiler choked on this
2355 * being argument to DEBUG_r(), so let's just have
2358 #ifdef PERL_EXT_RE_BUILD
2364 regnode *fix = convert;
2365 U32 word = trie->wordcount;
2367 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2368 while( ++fix < n ) {
2369 Set_Node_Offset_Length(fix, 0, 0);
2372 SV ** const tmp = av_fetch( trie_words, word, 0 );
2374 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2375 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2377 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2385 NEXT_OFF(convert) = (U16)(tail - convert);
2386 DEBUG_r(optimize= n);
2392 if ( trie->maxlen ) {
2393 NEXT_OFF( convert ) = (U16)(tail - convert);
2394 ARG_SET( convert, data_slot );
2395 /* Store the offset to the first unabsorbed branch in
2396 jump[0], which is otherwise unused by the jump logic.
2397 We use this when dumping a trie and during optimisation. */
2399 trie->jump[0] = (U16)(nextbranch - convert);
2401 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2402 * and there is a bitmap
2403 * and the first "jump target" node we found leaves enough room
2404 * then convert the TRIE node into a TRIEC node, with the bitmap
2405 * embedded inline in the opcode - this is hypothetically faster.
2407 if ( !trie->states[trie->startstate].wordnum
2409 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2411 OP( convert ) = TRIEC;
2412 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2413 PerlMemShared_free(trie->bitmap);
2416 OP( convert ) = TRIE;
2418 /* store the type in the flags */
2419 convert->flags = nodetype;
2423 + regarglen[ OP( convert ) ];
2425 /* XXX We really should free up the resource in trie now,
2426 as we won't use them - (which resources?) dmq */
2428 /* needed for dumping*/
2429 DEBUG_r(if (optimize) {
2430 regnode *opt = convert;
2432 while ( ++opt < optimize) {
2433 Set_Node_Offset_Length(opt,0,0);
2436 Try to clean up some of the debris left after the
2439 while( optimize < jumper ) {
2440 mjd_nodelen += Node_Length((optimize));
2441 OP( optimize ) = OPTIMIZED;
2442 Set_Node_Offset_Length(optimize,0,0);
2445 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2447 } /* end node insert */
2449 /* Finish populating the prev field of the wordinfo array. Walk back
2450 * from each accept state until we find another accept state, and if
2451 * so, point the first word's .prev field at the second word. If the
2452 * second already has a .prev field set, stop now. This will be the
2453 * case either if we've already processed that word's accept state,
2454 * or that state had multiple words, and the overspill words were
2455 * already linked up earlier.
2462 for (word=1; word <= trie->wordcount; word++) {
2464 if (trie->wordinfo[word].prev)
2466 state = trie->wordinfo[word].accept;
2468 state = prev_states[state];
2471 prev = trie->states[state].wordnum;
2475 trie->wordinfo[word].prev = prev;
2477 Safefree(prev_states);
2481 /* and now dump out the compressed format */
2482 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2484 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2486 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2487 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2489 SvREFCNT_dec_NN(revcharmap);
2493 : trie->startstate>1
2499 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2501 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2503 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2504 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2507 We find the fail state for each state in the trie, this state is the longest proper
2508 suffix of the current state's 'word' that is also a proper prefix of another word in our
2509 trie. State 1 represents the word '' and is thus the default fail state. This allows
2510 the DFA not to have to restart after its tried and failed a word at a given point, it
2511 simply continues as though it had been matching the other word in the first place.
2513 'abcdgu'=~/abcdefg|cdgu/
2514 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2515 fail, which would bring us to the state representing 'd' in the second word where we would
2516 try 'g' and succeed, proceeding to match 'cdgu'.
2518 /* add a fail transition */
2519 const U32 trie_offset = ARG(source);
2520 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2522 const U32 ucharcount = trie->uniquecharcount;
2523 const U32 numstates = trie->statecount;
2524 const U32 ubound = trie->lasttrans + ucharcount;
2528 U32 base = trie->states[ 1 ].trans.base;
2531 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2532 GET_RE_DEBUG_FLAGS_DECL;
2534 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2536 PERL_UNUSED_ARG(depth);
2540 ARG_SET( stclass, data_slot );
2541 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2542 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2543 aho->trie=trie_offset;
2544 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2545 Copy( trie->states, aho->states, numstates, reg_trie_state );
2546 Newxz( q, numstates, U32);
2547 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2550 /* initialize fail[0..1] to be 1 so that we always have
2551 a valid final fail state */
2552 fail[ 0 ] = fail[ 1 ] = 1;
2554 for ( charid = 0; charid < ucharcount ; charid++ ) {
2555 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2557 q[ q_write ] = newstate;
2558 /* set to point at the root */
2559 fail[ q[ q_write++ ] ]=1;
2562 while ( q_read < q_write) {
2563 const U32 cur = q[ q_read++ % numstates ];
2564 base = trie->states[ cur ].trans.base;
2566 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2567 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2569 U32 fail_state = cur;
2572 fail_state = fail[ fail_state ];
2573 fail_base = aho->states[ fail_state ].trans.base;
2574 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2576 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2577 fail[ ch_state ] = fail_state;
2578 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2580 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2582 q[ q_write++ % numstates] = ch_state;
2586 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2587 when we fail in state 1, this allows us to use the
2588 charclass scan to find a valid start char. This is based on the principle
2589 that theres a good chance the string being searched contains lots of stuff
2590 that cant be a start char.
2592 fail[ 0 ] = fail[ 1 ] = 0;
2593 DEBUG_TRIE_COMPILE_r({
2594 PerlIO_printf(Perl_debug_log,
2595 "%*sStclass Failtable (%"UVuf" states): 0",
2596 (int)(depth * 2), "", (UV)numstates
2598 for( q_read=1; q_read<numstates; q_read++ ) {
2599 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2601 PerlIO_printf(Perl_debug_log, "\n");
2604 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2609 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2610 * These need to be revisited when a newer toolchain becomes available.
2612 #if defined(__sparc64__) && defined(__GNUC__)
2613 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2614 # undef SPARC64_GCC_WORKAROUND
2615 # define SPARC64_GCC_WORKAROUND 1
2619 #define DEBUG_PEEP(str,scan,depth) \
2620 DEBUG_OPTIMISE_r({if (scan){ \
2621 SV * const mysv=sv_newmortal(); \
2622 regnode *Next = regnext(scan); \
2623 regprop(RExC_rx, mysv, scan); \
2624 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2625 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2626 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2630 /* The below joins as many adjacent EXACTish nodes as possible into a single
2631 * one. The regop may be changed if the node(s) contain certain sequences that
2632 * require special handling. The joining is only done if:
2633 * 1) there is room in the current conglomerated node to entirely contain the
2635 * 2) they are the exact same node type
2637 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2638 * these get optimized out
2640 * If a node is to match under /i (folded), the number of characters it matches
2641 * can be different than its character length if it contains a multi-character
2642 * fold. *min_subtract is set to the total delta of the input nodes.
2644 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2645 * and contains LATIN SMALL LETTER SHARP S
2647 * This is as good a place as any to discuss the design of handling these
2648 * multi-character fold sequences. It's been wrong in Perl for a very long
2649 * time. There are three code points in Unicode whose multi-character folds
2650 * were long ago discovered to mess things up. The previous designs for
2651 * dealing with these involved assigning a special node for them. This
2652 * approach doesn't work, as evidenced by this example:
2653 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2654 * Both these fold to "sss", but if the pattern is parsed to create a node that
2655 * would match just the \xDF, it won't be able to handle the case where a
2656 * successful match would have to cross the node's boundary. The new approach
2657 * that hopefully generally solves the problem generates an EXACTFU_SS node
2660 * It turns out that there are problems with all multi-character folds, and not
2661 * just these three. Now the code is general, for all such cases, but the
2662 * three still have some special handling. The approach taken is:
2663 * 1) This routine examines each EXACTFish node that could contain multi-
2664 * character fold sequences. It returns in *min_subtract how much to
2665 * subtract from the the actual length of the string to get a real minimum
2666 * match length; it is 0 if there are no multi-char folds. This delta is
2667 * used by the caller to adjust the min length of the match, and the delta
2668 * between min and max, so that the optimizer doesn't reject these
2669 * possibilities based on size constraints.
2670 * 2) Certain of these sequences require special handling by the trie code,
2671 * so, if found, this code changes the joined node type to special ops:
2672 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2673 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2674 * is used for an EXACTFU node that contains at least one "ss" sequence in
2675 * it. For non-UTF-8 patterns and strings, this is the only case where
2676 * there is a possible fold length change. That means that a regular
2677 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2678 * with length changes, and so can be processed faster. regexec.c takes
2679 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2680 * pre-folded by regcomp.c. This saves effort in regex matching.
2681 * However, the pre-folding isn't done for non-UTF8 patterns because the
2682 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2683 * down by forcing the pattern into UTF8 unless necessary. Also what
2684 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2685 * possibilities for the non-UTF8 patterns are quite simple, except for
2686 * the sharp s. All the ones that don't involve a UTF-8 target string are
2687 * members of a fold-pair, and arrays are set up for all of them so that
2688 * the other member of the pair can be found quickly. Code elsewhere in
2689 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2690 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2691 * described in the next item.
2692 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2693 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2694 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2695 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2696 * character in the pattern corresponds to at most a single character in
2697 * the target string. (And I do mean character, and not byte here, unlike
2698 * other parts of the documentation that have never been updated to
2699 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2700 * two character string 'ss'; in EXACTFA nodes it can match
2701 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2702 * instances where it is violated. I'm reluctant to try to change the
2703 * assumption, as the code involved is impenetrable to me (khw), so
2704 * instead the code here punts. This routine examines (when the pattern
2705 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2706 * boolean indicating whether or not the node contains a sharp s. When it
2707 * is true, the caller sets a flag that later causes the optimizer in this
2708 * file to not set values for the floating and fixed string lengths, and
2709 * thus avoids the optimizer code in regexec.c that makes the invalid
2710 * assumption. Thus, there is no optimization based on string lengths for
2711 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2712 * (The reason the assumption is wrong only in these two cases is that all
2713 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2714 * other folds to their expanded versions. We can't prefold sharp s to
2715 * 'ss' in EXACTF nodes because we don't know at compile time if it
2716 * actually matches 'ss' or not. It will match iff the target string is
2717 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2718 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2719 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2720 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2721 * require the pattern to be forced into UTF-8, the overhead of which we
2725 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2726 if (PL_regkind[OP(scan)] == EXACT) \
2727 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2730 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, UV *min_subtract, bool *has_exactf_sharp_s, U32 flags,regnode *val, U32 depth) {
2731 /* Merge several consecutive EXACTish nodes into one. */
2732 regnode *n = regnext(scan);
2734 regnode *next = scan + NODE_SZ_STR(scan);
2738 regnode *stop = scan;
2739 GET_RE_DEBUG_FLAGS_DECL;
2741 PERL_UNUSED_ARG(depth);
2744 PERL_ARGS_ASSERT_JOIN_EXACT;
2745 #ifndef EXPERIMENTAL_INPLACESCAN
2746 PERL_UNUSED_ARG(flags);
2747 PERL_UNUSED_ARG(val);
2749 DEBUG_PEEP("join",scan,depth);
2751 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2752 * EXACT ones that are mergeable to the current one. */
2754 && (PL_regkind[OP(n)] == NOTHING
2755 || (stringok && OP(n) == OP(scan)))
2757 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2760 if (OP(n) == TAIL || n > next)
2762 if (PL_regkind[OP(n)] == NOTHING) {
2763 DEBUG_PEEP("skip:",n,depth);
2764 NEXT_OFF(scan) += NEXT_OFF(n);
2765 next = n + NODE_STEP_REGNODE;
2772 else if (stringok) {
2773 const unsigned int oldl = STR_LEN(scan);
2774 regnode * const nnext = regnext(n);
2776 /* XXX I (khw) kind of doubt that this works on platforms where
2777 * U8_MAX is above 255 because of lots of other assumptions */
2778 /* Don't join if the sum can't fit into a single node */
2779 if (oldl + STR_LEN(n) > U8_MAX)
2782 DEBUG_PEEP("merg",n,depth);
2785 NEXT_OFF(scan) += NEXT_OFF(n);
2786 STR_LEN(scan) += STR_LEN(n);
2787 next = n + NODE_SZ_STR(n);
2788 /* Now we can overwrite *n : */
2789 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2797 #ifdef EXPERIMENTAL_INPLACESCAN
2798 if (flags && !NEXT_OFF(n)) {
2799 DEBUG_PEEP("atch", val, depth);
2800 if (reg_off_by_arg[OP(n)]) {
2801 ARG_SET(n, val - n);
2804 NEXT_OFF(n) = val - n;
2812 *has_exactf_sharp_s = FALSE;
2814 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2815 * can now analyze for sequences of problematic code points. (Prior to
2816 * this final joining, sequences could have been split over boundaries, and
2817 * hence missed). The sequences only happen in folding, hence for any
2818 * non-EXACT EXACTish node */
2819 if (OP(scan) != EXACT) {
2820 const U8 * const s0 = (U8*) STRING(scan);
2822 const U8 * const s_end = s0 + STR_LEN(scan);
2824 /* One pass is made over the node's string looking for all the
2825 * possibilities. to avoid some tests in the loop, there are two main
2826 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2830 /* Examine the string for a multi-character fold sequence. UTF-8
2831 * patterns have all characters pre-folded by the time this code is
2833 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2834 length sequence we are looking for is 2 */
2837 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2838 if (! len) { /* Not a multi-char fold: get next char */
2843 /* Nodes with 'ss' require special handling, except for EXACTFL
2844 * and EXACTFA for which there is no multi-char fold to this */
2845 if (len == 2 && *s == 's' && *(s+1) == 's'
2846 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2849 OP(scan) = EXACTFU_SS;
2852 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2854 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2855 COMBINING_DIAERESIS_UTF8
2856 COMBINING_ACUTE_ACCENT_UTF8,
2858 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2859 COMBINING_DIAERESIS_UTF8
2860 COMBINING_ACUTE_ACCENT_UTF8,
2865 /* These two folds require special handling by trie's, so
2866 * change the node type to indicate this. If EXACTFA and
2867 * EXACTFL were ever to be handled by trie's, this would
2868 * have to be changed. If this node has already been
2869 * changed to EXACTFU_SS in this loop, leave it as is. (I
2870 * (khw) think it doesn't matter in regexec.c for UTF
2871 * patterns, but no need to change it */
2872 if (OP(scan) == EXACTFU) {
2873 OP(scan) = EXACTFU_TRICKYFOLD;
2877 else { /* Here is a generic multi-char fold. */
2878 const U8* multi_end = s + len;
2880 /* Count how many characters in it. In the case of /l and
2881 * /aa, no folds which contain ASCII code points are
2882 * allowed, so check for those, and skip if found. (In
2883 * EXACTFL, no folds are allowed to any Latin1 code point,
2884 * not just ASCII. But there aren't any of these
2885 * currently, nor ever likely, so don't take the time to
2886 * test for them. The code that generates the
2887 * is_MULTI_foo() macros croaks should one actually get put
2888 * into Unicode .) */
2889 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2890 count = utf8_length(s, multi_end);
2894 while (s < multi_end) {
2897 goto next_iteration;
2907 /* The delta is how long the sequence is minus 1 (1 is how long
2908 * the character that folds to the sequence is) */
2909 *min_subtract += count - 1;
2913 else if (OP(scan) == EXACTFA) {
2915 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2916 * fold to the ASCII range (and there are no existing ones in the
2917 * upper latin1 range). But, as outlined in the comments preceding
2918 * this function, we need to flag any occurrences of the sharp s */
2920 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2921 *has_exactf_sharp_s = TRUE;
2928 else if (OP(scan) != EXACTFL) {
2930 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2931 * multi-char folds that are all Latin1. (This code knows that
2932 * there are no current multi-char folds possible with EXACTFL,
2933 * relying on fold_grind.t to catch any errors if the very unlikely
2934 * event happens that some get added in future Unicode versions.)
2935 * As explained in the comments preceding this function, we look
2936 * also for the sharp s in EXACTF nodes; it can be in the final
2937 * position. Otherwise we can stop looking 1 byte earlier because
2938 * have to find at least two characters for a multi-fold */
2939 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2941 /* The below is perhaps overboard, but this allows us to save a
2942 * test each time through the loop at the expense of a mask. This
2943 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2944 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2945 * are 64. This uses an exclusive 'or' to find that bit and then
2946 * inverts it to form a mask, with just a single 0, in the bit
2947 * position where 'S' and 's' differ. */
2948 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2949 const U8 s_masked = 's' & S_or_s_mask;
2952 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2953 if (! len) { /* Not a multi-char fold. */
2954 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2956 *has_exactf_sharp_s = TRUE;
2963 && ((*s & S_or_s_mask) == s_masked)
2964 && ((*(s+1) & S_or_s_mask) == s_masked))
2967 /* EXACTF nodes need to know that the minimum length
2968 * changed so that a sharp s in the string can match this
2969 * ss in the pattern, but they remain EXACTF nodes, as they
2970 * won't match this unless the target string is is UTF-8,
2971 * which we don't know until runtime */
2972 if (OP(scan) != EXACTF) {
2973 OP(scan) = EXACTFU_SS;
2977 *min_subtract += len - 1;
2984 /* Allow dumping but overwriting the collection of skipped
2985 * ops and/or strings with fake optimized ops */
2986 n = scan + NODE_SZ_STR(scan);
2994 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2998 /* REx optimizer. Converts nodes into quicker variants "in place".
2999 Finds fixed substrings. */
3001 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3002 to the position after last scanned or to NULL. */
3004 #define INIT_AND_WITHP \
3005 assert(!and_withp); \
3006 Newx(and_withp,1,struct regnode_charclass_class); \
3007 SAVEFREEPV(and_withp)
3009 /* this is a chain of data about sub patterns we are processing that
3010 need to be handled separately/specially in study_chunk. Its so
3011 we can simulate recursion without losing state. */
3013 typedef struct scan_frame {
3014 regnode *last; /* last node to process in this frame */
3015 regnode *next; /* next node to process when last is reached */
3016 struct scan_frame *prev; /*previous frame*/
3017 I32 stop; /* what stopparen do we use */
3021 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3024 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3025 I32 *minlenp, I32 *deltap,
3030 struct regnode_charclass_class *and_withp,
3031 U32 flags, U32 depth)
3032 /* scanp: Start here (read-write). */
3033 /* deltap: Write maxlen-minlen here. */
3034 /* last: Stop before this one. */
3035 /* data: string data about the pattern */
3036 /* stopparen: treat close N as END */
3037 /* recursed: which subroutines have we recursed into */
3038 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3041 I32 min = 0; /* There must be at least this number of characters to match */
3043 regnode *scan = *scanp, *next;
3045 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3046 int is_inf_internal = 0; /* The studied chunk is infinite */
3047 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3048 scan_data_t data_fake;
3049 SV *re_trie_maxbuff = NULL;
3050 regnode *first_non_open = scan;
3051 I32 stopmin = I32_MAX;
3052 scan_frame *frame = NULL;
3053 GET_RE_DEBUG_FLAGS_DECL;
3055 PERL_ARGS_ASSERT_STUDY_CHUNK;
3058 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3062 while (first_non_open && OP(first_non_open) == OPEN)
3063 first_non_open=regnext(first_non_open);
3068 while ( scan && OP(scan) != END && scan < last ){
3069 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3070 node length to get a real minimum (because
3071 the folded version may be shorter) */
3072 bool has_exactf_sharp_s = FALSE;
3073 /* Peephole optimizer: */
3074 DEBUG_STUDYDATA("Peep:", data,depth);
3075 DEBUG_PEEP("Peep",scan,depth);
3077 /* Its not clear to khw or hv why this is done here, and not in the
3078 * clauses that deal with EXACT nodes. khw's guess is that it's
3079 * because of a previous design */
3080 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3082 /* Follow the next-chain of the current node and optimize
3083 away all the NOTHINGs from it. */
3084 if (OP(scan) != CURLYX) {
3085 const int max = (reg_off_by_arg[OP(scan)]
3087 /* I32 may be smaller than U16 on CRAYs! */
3088 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3089 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3093 /* Skip NOTHING and LONGJMP. */
3094 while ((n = regnext(n))
3095 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3096 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3097 && off + noff < max)
3099 if (reg_off_by_arg[OP(scan)])
3102 NEXT_OFF(scan) = off;
3107 /* The principal pseudo-switch. Cannot be a switch, since we
3108 look into several different things. */
3109 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3110 || OP(scan) == IFTHEN) {
3111 next = regnext(scan);
3113 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3115 if (OP(next) == code || code == IFTHEN) {
3116 /* NOTE - There is similar code to this block below for handling
3117 TRIE nodes on a re-study. If you change stuff here check there
3119 I32 max1 = 0, min1 = I32_MAX, num = 0;
3120 struct regnode_charclass_class accum;
3121 regnode * const startbranch=scan;
3123 if (flags & SCF_DO_SUBSTR)
3124 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3125 if (flags & SCF_DO_STCLASS)
3126 cl_init_zero(pRExC_state, &accum);
3128 while (OP(scan) == code) {
3129 I32 deltanext, minnext, f = 0, fake;
3130 struct regnode_charclass_class this_class;
3133 data_fake.flags = 0;
3135 data_fake.whilem_c = data->whilem_c;
3136 data_fake.last_closep = data->last_closep;
3139 data_fake.last_closep = &fake;
3141 data_fake.pos_delta = delta;
3142 next = regnext(scan);
3143 scan = NEXTOPER(scan);
3145 scan = NEXTOPER(scan);
3146 if (flags & SCF_DO_STCLASS) {
3147 cl_init(pRExC_state, &this_class);
3148 data_fake.start_class = &this_class;
3149 f = SCF_DO_STCLASS_AND;
3151 if (flags & SCF_WHILEM_VISITED_POS)
3152 f |= SCF_WHILEM_VISITED_POS;
3154 /* we suppose the run is continuous, last=next...*/
3155 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3157 stopparen, recursed, NULL, f,depth+1);
3160 if (deltanext == I32_MAX) {
3161 is_inf = is_inf_internal = 1;
3163 } else if (max1 < minnext + deltanext)
3164 max1 = minnext + deltanext;
3166 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3168 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3169 if ( stopmin > minnext)
3170 stopmin = min + min1;
3171 flags &= ~SCF_DO_SUBSTR;
3173 data->flags |= SCF_SEEN_ACCEPT;
3176 if (data_fake.flags & SF_HAS_EVAL)
3177 data->flags |= SF_HAS_EVAL;
3178 data->whilem_c = data_fake.whilem_c;
3180 if (flags & SCF_DO_STCLASS)
3181 cl_or(pRExC_state, &accum, &this_class);
3183 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3185 if (flags & SCF_DO_SUBSTR) {
3186 data->pos_min += min1;
3187 if (data->pos_delta >= I32_MAX - (max1 - min1))
3188 data->pos_delta = I32_MAX;
3190 data->pos_delta += max1 - min1;
3191 if (max1 != min1 || is_inf)
3192 data->longest = &(data->longest_float);
3195 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3198 delta += max1 - min1;
3199 if (flags & SCF_DO_STCLASS_OR) {
3200 cl_or(pRExC_state, data->start_class, &accum);
3202 cl_and(data->start_class, and_withp);
3203 flags &= ~SCF_DO_STCLASS;
3206 else if (flags & SCF_DO_STCLASS_AND) {
3208 cl_and(data->start_class, &accum);
3209 flags &= ~SCF_DO_STCLASS;
3212 /* Switch to OR mode: cache the old value of
3213 * data->start_class */
3215 StructCopy(data->start_class, and_withp,
3216 struct regnode_charclass_class);
3217 flags &= ~SCF_DO_STCLASS_AND;
3218 StructCopy(&accum, data->start_class,
3219 struct regnode_charclass_class);
3220 flags |= SCF_DO_STCLASS_OR;
3221 SET_SSC_EOS(data->start_class);
3225 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3228 Assuming this was/is a branch we are dealing with: 'scan' now
3229 points at the item that follows the branch sequence, whatever
3230 it is. We now start at the beginning of the sequence and look
3237 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3239 If we can find such a subsequence we need to turn the first
3240 element into a trie and then add the subsequent branch exact
3241 strings to the trie.
3245 1. patterns where the whole set of branches can be converted.
3247 2. patterns where only a subset can be converted.
3249 In case 1 we can replace the whole set with a single regop
3250 for the trie. In case 2 we need to keep the start and end
3253 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3254 becomes BRANCH TRIE; BRANCH X;
3256 There is an additional case, that being where there is a
3257 common prefix, which gets split out into an EXACT like node
3258 preceding the TRIE node.
3260 If x(1..n)==tail then we can do a simple trie, if not we make
3261 a "jump" trie, such that when we match the appropriate word
3262 we "jump" to the appropriate tail node. Essentially we turn
3263 a nested if into a case structure of sorts.
3268 if (!re_trie_maxbuff) {
3269 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3270 if (!SvIOK(re_trie_maxbuff))
3271 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3273 if ( SvIV(re_trie_maxbuff)>=0 ) {
3275 regnode *first = (regnode *)NULL;
3276 regnode *last = (regnode *)NULL;
3277 regnode *tail = scan;
3282 SV * const mysv = sv_newmortal(); /* for dumping */
3284 /* var tail is used because there may be a TAIL
3285 regop in the way. Ie, the exacts will point to the
3286 thing following the TAIL, but the last branch will
3287 point at the TAIL. So we advance tail. If we
3288 have nested (?:) we may have to move through several
3292 while ( OP( tail ) == TAIL ) {
3293 /* this is the TAIL generated by (?:) */
3294 tail = regnext( tail );
3298 DEBUG_TRIE_COMPILE_r({
3299 regprop(RExC_rx, mysv, tail );
3300 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3301 (int)depth * 2 + 2, "",
3302 "Looking for TRIE'able sequences. Tail node is: ",
3303 SvPV_nolen_const( mysv )
3309 Step through the branches
3310 cur represents each branch,
3311 noper is the first thing to be matched as part of that branch
3312 noper_next is the regnext() of that node.
3314 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3315 via a "jump trie" but we also support building with NOJUMPTRIE,
3316 which restricts the trie logic to structures like /FOO|BAR/.
3318 If noper is a trieable nodetype then the branch is a possible optimization
3319 target. If we are building under NOJUMPTRIE then we require that noper_next
3320 is the same as scan (our current position in the regex program).
3322 Once we have two or more consecutive such branches we can create a
3323 trie of the EXACT's contents and stitch it in place into the program.
3325 If the sequence represents all of the branches in the alternation we
3326 replace the entire thing with a single TRIE node.
3328 Otherwise when it is a subsequence we need to stitch it in place and
3329 replace only the relevant branches. This means the first branch has
3330 to remain as it is used by the alternation logic, and its next pointer,
3331 and needs to be repointed at the item on the branch chain following
3332 the last branch we have optimized away.
3334 This could be either a BRANCH, in which case the subsequence is internal,
3335 or it could be the item following the branch sequence in which case the
3336 subsequence is at the end (which does not necessarily mean the first node
3337 is the start of the alternation).
3339 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3342 ----------------+-----------
3346 EXACTFU_SS | EXACTFU
3347 EXACTFU_TRICKYFOLD | EXACTFU
3352 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3353 ( EXACT == (X) ) ? EXACT : \
3354 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3357 /* dont use tail as the end marker for this traverse */
3358 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3359 regnode * const noper = NEXTOPER( cur );
3360 U8 noper_type = OP( noper );
3361 U8 noper_trietype = TRIE_TYPE( noper_type );
3362 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3363 regnode * const noper_next = regnext( noper );
3364 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3365 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3368 DEBUG_TRIE_COMPILE_r({
3369 regprop(RExC_rx, mysv, cur);
3370 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3371 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3373 regprop(RExC_rx, mysv, noper);
3374 PerlIO_printf( Perl_debug_log, " -> %s",
3375 SvPV_nolen_const(mysv));
3378 regprop(RExC_rx, mysv, noper_next );
3379 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3380 SvPV_nolen_const(mysv));
3382 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3383 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3384 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3388 /* Is noper a trieable nodetype that can be merged with the
3389 * current trie (if there is one)? */
3393 ( noper_trietype == NOTHING)
3394 || ( trietype == NOTHING )
3395 || ( trietype == noper_trietype )
3398 && noper_next == tail
3402 /* Handle mergable triable node
3403 * Either we are the first node in a new trieable sequence,
3404 * in which case we do some bookkeeping, otherwise we update
3405 * the end pointer. */
3408 if ( noper_trietype == NOTHING ) {
3409 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3410 regnode * const noper_next = regnext( noper );
3411 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3412 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3415 if ( noper_next_trietype ) {
3416 trietype = noper_next_trietype;
3417 } else if (noper_next_type) {
3418 /* a NOTHING regop is 1 regop wide. We need at least two
3419 * for a trie so we can't merge this in */
3423 trietype = noper_trietype;
3426 if ( trietype == NOTHING )
3427 trietype = noper_trietype;
3432 } /* end handle mergable triable node */
3434 /* handle unmergable node -
3435 * noper may either be a triable node which can not be tried
3436 * together with the current trie, or a non triable node */
3438 /* If last is set and trietype is not NOTHING then we have found
3439 * at least two triable branch sequences in a row of a similar
3440 * trietype so we can turn them into a trie. If/when we
3441 * allow NOTHING to start a trie sequence this condition will be
3442 * required, and it isn't expensive so we leave it in for now. */
3443 if ( trietype && trietype != NOTHING )
3444 make_trie( pRExC_state,
3445 startbranch, first, cur, tail, count,
3446 trietype, depth+1 );
3447 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3451 && noper_next == tail
3454 /* noper is triable, so we can start a new trie sequence */
3457 trietype = noper_trietype;
3459 /* if we already saw a first but the current node is not triable then we have
3460 * to reset the first information. */
3465 } /* end handle unmergable node */
3466 } /* loop over branches */
3467 DEBUG_TRIE_COMPILE_r({
3468 regprop(RExC_rx, mysv, cur);
3469 PerlIO_printf( Perl_debug_log,
3470 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3471 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3474 if ( last && trietype ) {
3475 if ( trietype != NOTHING ) {
3476 /* the last branch of the sequence was part of a trie,
3477 * so we have to construct it here outside of the loop
3479 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3480 #ifdef TRIE_STUDY_OPT
3481 if ( ((made == MADE_EXACT_TRIE &&
3482 startbranch == first)
3483 || ( first_non_open == first )) &&
3485 flags |= SCF_TRIE_RESTUDY;
3486 if ( startbranch == first
3489 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3494 /* at this point we know whatever we have is a NOTHING sequence/branch
3495 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3497 if ( startbranch == first ) {
3499 /* the entire thing is a NOTHING sequence, something like this:
3500 * (?:|) So we can turn it into a plain NOTHING op. */
3501 DEBUG_TRIE_COMPILE_r({
3502 regprop(RExC_rx, mysv, cur);
3503 PerlIO_printf( Perl_debug_log,
3504 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3505 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3508 OP(startbranch)= NOTHING;
3509 NEXT_OFF(startbranch)= tail - startbranch;
3510 for ( opt= startbranch + 1; opt < tail ; opt++ )
3514 } /* end if ( last) */
3515 } /* TRIE_MAXBUF is non zero */
3520 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3521 scan = NEXTOPER(NEXTOPER(scan));
3522 } else /* single branch is optimized. */
3523 scan = NEXTOPER(scan);
3525 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3526 scan_frame *newframe = NULL;
3531 if (OP(scan) != SUSPEND) {
3532 /* set the pointer */
3533 if (OP(scan) == GOSUB) {
3535 RExC_recurse[ARG2L(scan)] = scan;
3536 start = RExC_open_parens[paren-1];
3537 end = RExC_close_parens[paren-1];
3540 start = RExC_rxi->program + 1;
3544 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3545 SAVEFREEPV(recursed);
3547 if (!PAREN_TEST(recursed,paren+1)) {
3548 PAREN_SET(recursed,paren+1);
3549 Newx(newframe,1,scan_frame);
3551 if (flags & SCF_DO_SUBSTR) {
3552 SCAN_COMMIT(pRExC_state,data,minlenp);
3553 data->longest = &(data->longest_float);
3555 is_inf = is_inf_internal = 1;
3556 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3557 cl_anything(pRExC_state, data->start_class);
3558 flags &= ~SCF_DO_STCLASS;
3561 Newx(newframe,1,scan_frame);
3564 end = regnext(scan);
3569 SAVEFREEPV(newframe);
3570 newframe->next = regnext(scan);
3571 newframe->last = last;
3572 newframe->stop = stopparen;
3573 newframe->prev = frame;
3583 else if (OP(scan) == EXACT) {
3584 I32 l = STR_LEN(scan);
3587 const U8 * const s = (U8*)STRING(scan);
3588 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3589 l = utf8_length(s, s + l);
3591 uc = *((U8*)STRING(scan));
3594 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3595 /* The code below prefers earlier match for fixed
3596 offset, later match for variable offset. */
3597 if (data->last_end == -1) { /* Update the start info. */
3598 data->last_start_min = data->pos_min;
3599 data->last_start_max = is_inf
3600 ? I32_MAX : data->pos_min + data->pos_delta;
3602 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3604 SvUTF8_on(data->last_found);
3606 SV * const sv = data->last_found;
3607 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3608 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3609 if (mg && mg->mg_len >= 0)
3610 mg->mg_len += utf8_length((U8*)STRING(scan),
3611 (U8*)STRING(scan)+STR_LEN(scan));
3613 data->last_end = data->pos_min + l;
3614 data->pos_min += l; /* As in the first entry. */
3615 data->flags &= ~SF_BEFORE_EOL;
3617 if (flags & SCF_DO_STCLASS_AND) {
3618 /* Check whether it is compatible with what we know already! */
3622 /* If compatible, we or it in below. It is compatible if is
3623 * in the bitmp and either 1) its bit or its fold is set, or 2)
3624 * it's for a locale. Even if there isn't unicode semantics
3625 * here, at runtime there may be because of matching against a
3626 * utf8 string, so accept a possible false positive for
3627 * latin1-range folds */
3629 (!(data->start_class->flags & ANYOF_LOCALE)
3630 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3631 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3632 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3637 ANYOF_CLASS_ZERO(data->start_class);
3638 ANYOF_BITMAP_ZERO(data->start_class);
3640 ANYOF_BITMAP_SET(data->start_class, uc);
3641 else if (uc >= 0x100) {
3644 /* Some Unicode code points fold to the Latin1 range; as
3645 * XXX temporary code, instead of figuring out if this is
3646 * one, just assume it is and set all the start class bits
3647 * that could be some such above 255 code point's fold
3648 * which will generate fals positives. As the code
3649 * elsewhere that does compute the fold settles down, it
3650 * can be extracted out and re-used here */
3651 for (i = 0; i < 256; i++){
3652 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3653 ANYOF_BITMAP_SET(data->start_class, i);
3657 CLEAR_SSC_EOS(data->start_class);
3659 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3661 else if (flags & SCF_DO_STCLASS_OR) {
3662 /* false positive possible if the class is case-folded */
3664 ANYOF_BITMAP_SET(data->start_class, uc);
3666 data->start_class->flags |= ANYOF_UNICODE_ALL;
3667 CLEAR_SSC_EOS(data->start_class);
3668 cl_and(data->start_class, and_withp);
3670 flags &= ~SCF_DO_STCLASS;
3672 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3673 I32 l = STR_LEN(scan);
3674 UV uc = *((U8*)STRING(scan));
3676 /* Search for fixed substrings supports EXACT only. */
3677 if (flags & SCF_DO_SUBSTR) {
3679 SCAN_COMMIT(pRExC_state, data, minlenp);
3682 const U8 * const s = (U8 *)STRING(scan);
3683 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3684 l = utf8_length(s, s + l);
3686 if (has_exactf_sharp_s) {
3687 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3689 min += l - min_subtract;
3691 delta += min_subtract;
3692 if (flags & SCF_DO_SUBSTR) {
3693 data->pos_min += l - min_subtract;
3694 if (data->pos_min < 0) {
3697 data->pos_delta += min_subtract;
3699 data->longest = &(data->longest_float);
3702 if (flags & SCF_DO_STCLASS_AND) {
3703 /* Check whether it is compatible with what we know already! */
3706 (!(data->start_class->flags & ANYOF_LOCALE)
3707 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3708 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3712 ANYOF_CLASS_ZERO(data->start_class);
3713 ANYOF_BITMAP_ZERO(data->start_class);
3715 ANYOF_BITMAP_SET(data->start_class, uc);
3716 CLEAR_SSC_EOS(data->start_class);
3717 if (OP(scan) == EXACTFL) {
3718 /* XXX This set is probably no longer necessary, and
3719 * probably wrong as LOCALE now is on in the initial
3721 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3725 /* Also set the other member of the fold pair. In case
3726 * that unicode semantics is called for at runtime, use
3727 * the full latin1 fold. (Can't do this for locale,
3728 * because not known until runtime) */
3729 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3731 /* All other (EXACTFL handled above) folds except under
3732 * /iaa that include s, S, and sharp_s also may include
3734 if (OP(scan) != EXACTFA) {
3735 if (uc == 's' || uc == 'S') {
3736 ANYOF_BITMAP_SET(data->start_class,
3737 LATIN_SMALL_LETTER_SHARP_S);
3739 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3740 ANYOF_BITMAP_SET(data->start_class, 's');
3741 ANYOF_BITMAP_SET(data->start_class, 'S');
3746 else if (uc >= 0x100) {
3748 for (i = 0; i < 256; i++){
3749 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3750 ANYOF_BITMAP_SET(data->start_class, i);
3755 else if (flags & SCF_DO_STCLASS_OR) {
3756 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3757 /* false positive possible if the class is case-folded.
3758 Assume that the locale settings are the same... */
3760 ANYOF_BITMAP_SET(data->start_class, uc);
3761 if (OP(scan) != EXACTFL) {
3763 /* And set the other member of the fold pair, but
3764 * can't do that in locale because not known until
3766 ANYOF_BITMAP_SET(data->start_class,
3767 PL_fold_latin1[uc]);
3769 /* All folds except under /iaa that include s, S,
3770 * and sharp_s also may include the others */
3771 if (OP(scan) != EXACTFA) {
3772 if (uc == 's' || uc == 'S') {
3773 ANYOF_BITMAP_SET(data->start_class,
3774 LATIN_SMALL_LETTER_SHARP_S);
3776 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3777 ANYOF_BITMAP_SET(data->start_class, 's');
3778 ANYOF_BITMAP_SET(data->start_class, 'S');
3783 CLEAR_SSC_EOS(data->start_class);
3785 cl_and(data->start_class, and_withp);
3787 flags &= ~SCF_DO_STCLASS;
3789 else if (REGNODE_VARIES(OP(scan))) {
3790 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3791 I32 f = flags, pos_before = 0;
3792 regnode * const oscan = scan;
3793 struct regnode_charclass_class this_class;
3794 struct regnode_charclass_class *oclass = NULL;
3795 I32 next_is_eval = 0;
3797 switch (PL_regkind[OP(scan)]) {
3798 case WHILEM: /* End of (?:...)* . */
3799 scan = NEXTOPER(scan);
3802 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3803 next = NEXTOPER(scan);
3804 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3806 maxcount = REG_INFTY;
3807 next = regnext(scan);
3808 scan = NEXTOPER(scan);
3812 if (flags & SCF_DO_SUBSTR)
3817 if (flags & SCF_DO_STCLASS) {
3819 maxcount = REG_INFTY;
3820 next = regnext(scan);
3821 scan = NEXTOPER(scan);
3824 is_inf = is_inf_internal = 1;
3825 scan = regnext(scan);
3826 if (flags & SCF_DO_SUBSTR) {
3827 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3828 data->longest = &(data->longest_float);
3830 goto optimize_curly_tail;
3832 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3833 && (scan->flags == stopparen))
3838 mincount = ARG1(scan);
3839 maxcount = ARG2(scan);
3841 next = regnext(scan);
3842 if (OP(scan) == CURLYX) {
3843 I32 lp = (data ? *(data->last_closep) : 0);
3844 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3846 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3847 next_is_eval = (OP(scan) == EVAL);
3849 if (flags & SCF_DO_SUBSTR) {
3850 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3851 pos_before = data->pos_min;
3855 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3857 data->flags |= SF_IS_INF;
3859 if (flags & SCF_DO_STCLASS) {
3860 cl_init(pRExC_state, &this_class);
3861 oclass = data->start_class;
3862 data->start_class = &this_class;
3863 f |= SCF_DO_STCLASS_AND;
3864 f &= ~SCF_DO_STCLASS_OR;
3866 /* Exclude from super-linear cache processing any {n,m}
3867 regops for which the combination of input pos and regex
3868 pos is not enough information to determine if a match
3871 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3872 regex pos at the \s*, the prospects for a match depend not
3873 only on the input position but also on how many (bar\s*)
3874 repeats into the {4,8} we are. */
3875 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3876 f &= ~SCF_WHILEM_VISITED_POS;
3878 /* This will finish on WHILEM, setting scan, or on NULL: */
3879 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3880 last, data, stopparen, recursed, NULL,
3882 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3884 if (flags & SCF_DO_STCLASS)
3885 data->start_class = oclass;
3886 if (mincount == 0 || minnext == 0) {
3887 if (flags & SCF_DO_STCLASS_OR) {
3888 cl_or(pRExC_state, data->start_class, &this_class);
3890 else if (flags & SCF_DO_STCLASS_AND) {
3891 /* Switch to OR mode: cache the old value of
3892 * data->start_class */
3894 StructCopy(data->start_class, and_withp,
3895 struct regnode_charclass_class);
3896 flags &= ~SCF_DO_STCLASS_AND;
3897 StructCopy(&this_class, data->start_class,
3898 struct regnode_charclass_class);
3899 flags |= SCF_DO_STCLASS_OR;
3900 SET_SSC_EOS(data->start_class);
3902 } else { /* Non-zero len */
3903 if (flags & SCF_DO_STCLASS_OR) {
3904 cl_or(pRExC_state, data->start_class, &this_class);
3905 cl_and(data->start_class, and_withp);
3907 else if (flags & SCF_DO_STCLASS_AND)
3908 cl_and(data->start_class, &this_class);
3909 flags &= ~SCF_DO_STCLASS;
3911 if (!scan) /* It was not CURLYX, but CURLY. */
3913 if (!(flags & SCF_TRIE_DOING_RESTUDY)
3914 /* ? quantifier ok, except for (?{ ... }) */
3915 && (next_is_eval || !(mincount == 0 && maxcount == 1))
3916 && (minnext == 0) && (deltanext == 0)
3917 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3918 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3920 /* Fatal warnings may leak the regexp without this: */
3921 SAVEFREESV(RExC_rx_sv);
3922 ckWARNreg(RExC_parse,
3923 "Quantifier unexpected on zero-length expression");
3924 (void)ReREFCNT_inc(RExC_rx_sv);
3927 min += minnext * mincount;
3928 is_inf_internal |= deltanext == I32_MAX
3929 || (maxcount == REG_INFTY && minnext + deltanext > 0);
3930 is_inf |= is_inf_internal;
3934 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3936 /* Try powerful optimization CURLYX => CURLYN. */
3937 if ( OP(oscan) == CURLYX && data
3938 && data->flags & SF_IN_PAR
3939 && !(data->flags & SF_HAS_EVAL)
3940 && !deltanext && minnext == 1 ) {
3941 /* Try to optimize to CURLYN. */
3942 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3943 regnode * const nxt1 = nxt;
3950 if (!REGNODE_SIMPLE(OP(nxt))
3951 && !(PL_regkind[OP(nxt)] == EXACT
3952 && STR_LEN(nxt) == 1))
3958 if (OP(nxt) != CLOSE)
3960 if (RExC_open_parens) {
3961 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3962 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3964 /* Now we know that nxt2 is the only contents: */
3965 oscan->flags = (U8)ARG(nxt);
3967 OP(nxt1) = NOTHING; /* was OPEN. */
3970 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3971 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3972 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3973 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3974 OP(nxt + 1) = OPTIMIZED; /* was count. */
3975 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3980 /* Try optimization CURLYX => CURLYM. */
3981 if ( OP(oscan) == CURLYX && data
3982 && !(data->flags & SF_HAS_PAR)
3983 && !(data->flags & SF_HAS_EVAL)
3984 && !deltanext /* atom is fixed width */
3985 && minnext != 0 /* CURLYM can't handle zero width */
3986 && ! (RExC_seen & REG_SEEN_EXACTF_SHARP_S) /* Nor \xDF */
3988 /* XXXX How to optimize if data == 0? */
3989 /* Optimize to a simpler form. */
3990 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3994 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3995 && (OP(nxt2) != WHILEM))
3997 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3998 /* Need to optimize away parenths. */
3999 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4000 /* Set the parenth number. */
4001 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4003 oscan->flags = (U8)ARG(nxt);
4004 if (RExC_open_parens) {
4005 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4006 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4008 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4009 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4012 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4013 OP(nxt + 1) = OPTIMIZED; /* was count. */
4014 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4015 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4018 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4019 regnode *nnxt = regnext(nxt1);
4021 if (reg_off_by_arg[OP(nxt1)])
4022 ARG_SET(nxt1, nxt2 - nxt1);
4023 else if (nxt2 - nxt1 < U16_MAX)
4024 NEXT_OFF(nxt1) = nxt2 - nxt1;
4026 OP(nxt) = NOTHING; /* Cannot beautify */
4031 /* Optimize again: */
4032 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4033 NULL, stopparen, recursed, NULL, 0,depth+1);
4038 else if ((OP(oscan) == CURLYX)
4039 && (flags & SCF_WHILEM_VISITED_POS)
4040 /* See the comment on a similar expression above.
4041 However, this time it's not a subexpression
4042 we care about, but the expression itself. */
4043 && (maxcount == REG_INFTY)
4044 && data && ++data->whilem_c < 16) {
4045 /* This stays as CURLYX, we can put the count/of pair. */
4046 /* Find WHILEM (as in regexec.c) */
4047 regnode *nxt = oscan + NEXT_OFF(oscan);
4049 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4051 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4052 | (RExC_whilem_seen << 4)); /* On WHILEM */
4054 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4056 if (flags & SCF_DO_SUBSTR) {
4057 SV *last_str = NULL;
4058 int counted = mincount != 0;
4060 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
4061 #if defined(SPARC64_GCC_WORKAROUND)
4064 const char *s = NULL;
4067 if (pos_before >= data->last_start_min)
4070 b = data->last_start_min;
4073 s = SvPV_const(data->last_found, l);
4074 old = b - data->last_start_min;
4077 I32 b = pos_before >= data->last_start_min
4078 ? pos_before : data->last_start_min;
4080 const char * const s = SvPV_const(data->last_found, l);
4081 I32 old = b - data->last_start_min;
4085 old = utf8_hop((U8*)s, old) - (U8*)s;
4087 /* Get the added string: */
4088 last_str = newSVpvn_utf8(s + old, l, UTF);
4089 if (deltanext == 0 && pos_before == b) {
4090 /* What was added is a constant string */
4092 SvGROW(last_str, (mincount * l) + 1);
4093 repeatcpy(SvPVX(last_str) + l,
4094 SvPVX_const(last_str), l, mincount - 1);
4095 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4096 /* Add additional parts. */
4097 SvCUR_set(data->last_found,
4098 SvCUR(data->last_found) - l);
4099 sv_catsv(data->last_found, last_str);
4101 SV * sv = data->last_found;
4103 SvUTF8(sv) && SvMAGICAL(sv) ?
4104 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4105 if (mg && mg->mg_len >= 0)
4106 mg->mg_len += CHR_SVLEN(last_str) - l;
4108 data->last_end += l * (mincount - 1);
4111 /* start offset must point into the last copy */
4112 data->last_start_min += minnext * (mincount - 1);
4113 data->last_start_max += is_inf ? I32_MAX
4114 : (maxcount - 1) * (minnext + data->pos_delta);
4117 /* It is counted once already... */
4118 data->pos_min += minnext * (mincount - counted);
4120 PerlIO_printf(Perl_debug_log, "counted=%d deltanext=%d I32_MAX=%d minnext=%d maxcount=%d mincount=%d\n",
4121 counted, deltanext, I32_MAX, minnext, maxcount, mincount);
4122 if (deltanext != I32_MAX)
4123 PerlIO_printf(Perl_debug_log, "LHS=%d RHS=%d\n", -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount, I32_MAX - data->pos_delta);
4125 if (deltanext == I32_MAX || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= I32_MAX - data->pos_delta)
4126 data->pos_delta = I32_MAX;
4128 data->pos_delta += - counted * deltanext +
4129 (minnext + deltanext) * maxcount - minnext * mincount;
4130 if (mincount != maxcount) {
4131 /* Cannot extend fixed substrings found inside
4133 SCAN_COMMIT(pRExC_state,data,minlenp);
4134 if (mincount && last_str) {
4135 SV * const sv = data->last_found;
4136 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4137 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4141 sv_setsv(sv, last_str);
4142 data->last_end = data->pos_min;
4143 data->last_start_min =
4144 data->pos_min - CHR_SVLEN(last_str);
4145 data->last_start_max = is_inf
4147 : data->pos_min + data->pos_delta
4148 - CHR_SVLEN(last_str);
4150 data->longest = &(data->longest_float);
4152 SvREFCNT_dec(last_str);
4154 if (data && (fl & SF_HAS_EVAL))
4155 data->flags |= SF_HAS_EVAL;
4156 optimize_curly_tail:
4157 if (OP(oscan) != CURLYX) {
4158 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4160 NEXT_OFF(oscan) += NEXT_OFF(next);
4163 default: /* REF, and CLUMP only? */
4164 if (flags & SCF_DO_SUBSTR) {
4165 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4166 data->longest = &(data->longest_float);
4168 is_inf = is_inf_internal = 1;
4169 if (flags & SCF_DO_STCLASS_OR)
4170 cl_anything(pRExC_state, data->start_class);
4171 flags &= ~SCF_DO_STCLASS;
4175 else if (OP(scan) == LNBREAK) {
4176 if (flags & SCF_DO_STCLASS) {
4178 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4179 if (flags & SCF_DO_STCLASS_AND) {
4180 for (value = 0; value < 256; value++)
4181 if (!is_VERTWS_cp(value))
4182 ANYOF_BITMAP_CLEAR(data->start_class, value);
4185 for (value = 0; value < 256; value++)
4186 if (is_VERTWS_cp(value))
4187 ANYOF_BITMAP_SET(data->start_class, value);
4189 if (flags & SCF_DO_STCLASS_OR)
4190 cl_and(data->start_class, and_withp);
4191 flags &= ~SCF_DO_STCLASS;
4194 delta++; /* Because of the 2 char string cr-lf */
4195 if (flags & SCF_DO_SUBSTR) {
4196 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4198 data->pos_delta += 1;
4199 data->longest = &(data->longest_float);
4202 else if (REGNODE_SIMPLE(OP(scan))) {
4205 if (flags & SCF_DO_SUBSTR) {
4206 SCAN_COMMIT(pRExC_state,data,minlenp);
4210 if (flags & SCF_DO_STCLASS) {
4212 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4214 /* Some of the logic below assumes that switching
4215 locale on will only add false positives. */
4216 switch (PL_regkind[OP(scan)]) {
4222 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan));
4225 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4226 cl_anything(pRExC_state, data->start_class);
4229 if (OP(scan) == SANY)
4231 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
4232 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
4233 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
4234 cl_anything(pRExC_state, data->start_class);
4236 if (flags & SCF_DO_STCLASS_AND || !value)
4237 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
4240 if (flags & SCF_DO_STCLASS_AND)
4241 cl_and(data->start_class,
4242 (struct regnode_charclass_class*)scan);
4244 cl_or(pRExC_state, data->start_class,
4245 (struct regnode_charclass_class*)scan);
4253 classnum = FLAGS(scan);
4254 if (flags & SCF_DO_STCLASS_AND) {
4255 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4256 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum) + 1);
4257 for (value = 0; value < loop_max; value++) {
4258 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4259 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4265 if (data->start_class->flags & ANYOF_LOCALE) {
4266 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum));
4270 /* Even if under locale, set the bits for non-locale
4271 * in case it isn't a true locale-node. This will
4272 * create false positives if it truly is locale */
4273 for (value = 0; value < loop_max; value++) {
4274 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4275 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4287 classnum = FLAGS(scan);
4288 if (flags & SCF_DO_STCLASS_AND) {
4289 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4290 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum));
4291 for (value = 0; value < loop_max; value++) {
4292 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4293 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4299 if (data->start_class->flags & ANYOF_LOCALE) {
4300 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum) + 1);
4304 /* Even if under locale, set the bits for non-locale in
4305 * case it isn't a true locale-node. This will create
4306 * false positives if it truly is locale */
4307 for (value = 0; value < loop_max; value++) {
4308 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4309 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4312 if (PL_regkind[OP(scan)] == NPOSIXD) {
4313 data->start_class->flags |= ANYOF_NON_UTF8_LATIN1_ALL;
4319 if (flags & SCF_DO_STCLASS_OR)
4320 cl_and(data->start_class, and_withp);
4321 flags &= ~SCF_DO_STCLASS;
4324 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4325 data->flags |= (OP(scan) == MEOL
4328 SCAN_COMMIT(pRExC_state, data, minlenp);
4331 else if ( PL_regkind[OP(scan)] == BRANCHJ
4332 /* Lookbehind, or need to calculate parens/evals/stclass: */
4333 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4334 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4335 if ( OP(scan) == UNLESSM &&
4337 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4338 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4341 regnode *upto= regnext(scan);
4343 SV * const mysv_val=sv_newmortal();
4344 DEBUG_STUDYDATA("OPFAIL",data,depth);
4346 /*DEBUG_PARSE_MSG("opfail");*/
4347 regprop(RExC_rx, mysv_val, upto);
4348 PerlIO_printf(Perl_debug_log, "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4349 SvPV_nolen_const(mysv_val),
4350 (IV)REG_NODE_NUM(upto),
4355 NEXT_OFF(scan) = upto - scan;
4356 for (opt= scan + 1; opt < upto ; opt++)
4357 OP(opt) = OPTIMIZED;
4361 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4362 || OP(scan) == UNLESSM )
4364 /* Negative Lookahead/lookbehind
4365 In this case we can't do fixed string optimisation.
4368 I32 deltanext, minnext, fake = 0;
4370 struct regnode_charclass_class intrnl;
4373 data_fake.flags = 0;
4375 data_fake.whilem_c = data->whilem_c;
4376 data_fake.last_closep = data->last_closep;
4379 data_fake.last_closep = &fake;
4380 data_fake.pos_delta = delta;
4381 if ( flags & SCF_DO_STCLASS && !scan->flags
4382 && OP(scan) == IFMATCH ) { /* Lookahead */
4383 cl_init(pRExC_state, &intrnl);
4384 data_fake.start_class = &intrnl;
4385 f |= SCF_DO_STCLASS_AND;
4387 if (flags & SCF_WHILEM_VISITED_POS)
4388 f |= SCF_WHILEM_VISITED_POS;
4389 next = regnext(scan);
4390 nscan = NEXTOPER(NEXTOPER(scan));
4391 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4392 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
4395 FAIL("Variable length lookbehind not implemented");
4397 else if (minnext > (I32)U8_MAX) {
4398 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4400 scan->flags = (U8)minnext;
4403 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4405 if (data_fake.flags & SF_HAS_EVAL)
4406 data->flags |= SF_HAS_EVAL;
4407 data->whilem_c = data_fake.whilem_c;
4409 if (f & SCF_DO_STCLASS_AND) {
4410 if (flags & SCF_DO_STCLASS_OR) {
4411 /* OR before, AND after: ideally we would recurse with
4412 * data_fake to get the AND applied by study of the
4413 * remainder of the pattern, and then derecurse;
4414 * *** HACK *** for now just treat as "no information".
4415 * See [perl #56690].
4417 cl_init(pRExC_state, data->start_class);
4419 /* AND before and after: combine and continue */
4420 const int was = TEST_SSC_EOS(data->start_class);
4422 cl_and(data->start_class, &intrnl);
4424 SET_SSC_EOS(data->start_class);
4428 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4430 /* Positive Lookahead/lookbehind
4431 In this case we can do fixed string optimisation,
4432 but we must be careful about it. Note in the case of
4433 lookbehind the positions will be offset by the minimum
4434 length of the pattern, something we won't know about
4435 until after the recurse.
4437 I32 deltanext, fake = 0;
4439 struct regnode_charclass_class intrnl;
4441 /* We use SAVEFREEPV so that when the full compile
4442 is finished perl will clean up the allocated
4443 minlens when it's all done. This way we don't
4444 have to worry about freeing them when we know
4445 they wont be used, which would be a pain.
4448 Newx( minnextp, 1, I32 );
4449 SAVEFREEPV(minnextp);
4452 StructCopy(data, &data_fake, scan_data_t);
4453 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
4456 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
4457 data_fake.last_found=newSVsv(data->last_found);
4461 data_fake.last_closep = &fake;
4462 data_fake.flags = 0;
4463 data_fake.pos_delta = delta;
4465 data_fake.flags |= SF_IS_INF;
4466 if ( flags & SCF_DO_STCLASS && !scan->flags
4467 && OP(scan) == IFMATCH ) { /* Lookahead */
4468 cl_init(pRExC_state, &intrnl);
4469 data_fake.start_class = &intrnl;
4470 f |= SCF_DO_STCLASS_AND;
4472 if (flags & SCF_WHILEM_VISITED_POS)
4473 f |= SCF_WHILEM_VISITED_POS;
4474 next = regnext(scan);
4475 nscan = NEXTOPER(NEXTOPER(scan));
4477 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
4478 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
4481 FAIL("Variable length lookbehind not implemented");
4483 else if (*minnextp > (I32)U8_MAX) {
4484 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4486 scan->flags = (U8)*minnextp;
4491 if (f & SCF_DO_STCLASS_AND) {
4492 const int was = TEST_SSC_EOS(data.start_class);
4494 cl_and(data->start_class, &intrnl);
4496 SET_SSC_EOS(data->start_class);
4499 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4501 if (data_fake.flags & SF_HAS_EVAL)
4502 data->flags |= SF_HAS_EVAL;
4503 data->whilem_c = data_fake.whilem_c;
4504 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
4505 if (RExC_rx->minlen<*minnextp)
4506 RExC_rx->minlen=*minnextp;
4507 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
4508 SvREFCNT_dec_NN(data_fake.last_found);
4510 if ( data_fake.minlen_fixed != minlenp )
4512 data->offset_fixed= data_fake.offset_fixed;
4513 data->minlen_fixed= data_fake.minlen_fixed;
4514 data->lookbehind_fixed+= scan->flags;
4516 if ( data_fake.minlen_float != minlenp )
4518 data->minlen_float= data_fake.minlen_float;
4519 data->offset_float_min=data_fake.offset_float_min;
4520 data->offset_float_max=data_fake.offset_float_max;
4521 data->lookbehind_float+= scan->flags;
4528 else if (OP(scan) == OPEN) {
4529 if (stopparen != (I32)ARG(scan))
4532 else if (OP(scan) == CLOSE) {
4533 if (stopparen == (I32)ARG(scan)) {
4536 if ((I32)ARG(scan) == is_par) {
4537 next = regnext(scan);
4539 if ( next && (OP(next) != WHILEM) && next < last)
4540 is_par = 0; /* Disable optimization */
4543 *(data->last_closep) = ARG(scan);
4545 else if (OP(scan) == EVAL) {
4547 data->flags |= SF_HAS_EVAL;
4549 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
4550 if (flags & SCF_DO_SUBSTR) {
4551 SCAN_COMMIT(pRExC_state,data,minlenp);
4552 flags &= ~SCF_DO_SUBSTR;
4554 if (data && OP(scan)==ACCEPT) {
4555 data->flags |= SCF_SEEN_ACCEPT;
4560 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
4562 if (flags & SCF_DO_SUBSTR) {
4563 SCAN_COMMIT(pRExC_state,data,minlenp);
4564 data->longest = &(data->longest_float);
4566 is_inf = is_inf_internal = 1;
4567 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4568 cl_anything(pRExC_state, data->start_class);
4569 flags &= ~SCF_DO_STCLASS;
4571 else if (OP(scan) == GPOS) {
4572 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
4573 !(delta || is_inf || (data && data->pos_delta)))
4575 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
4576 RExC_rx->extflags |= RXf_ANCH_GPOS;
4577 if (RExC_rx->gofs < (U32)min)
4578 RExC_rx->gofs = min;
4580 RExC_rx->extflags |= RXf_GPOS_FLOAT;
4584 #ifdef TRIE_STUDY_OPT
4585 #ifdef FULL_TRIE_STUDY
4586 else if (PL_regkind[OP(scan)] == TRIE) {
4587 /* NOTE - There is similar code to this block above for handling
4588 BRANCH nodes on the initial study. If you change stuff here
4590 regnode *trie_node= scan;
4591 regnode *tail= regnext(scan);
4592 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4593 I32 max1 = 0, min1 = I32_MAX;
4594 struct regnode_charclass_class accum;
4596 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
4597 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
4598 if (flags & SCF_DO_STCLASS)
4599 cl_init_zero(pRExC_state, &accum);
4605 const regnode *nextbranch= NULL;
4608 for ( word=1 ; word <= trie->wordcount ; word++)
4610 I32 deltanext=0, minnext=0, f = 0, fake;
4611 struct regnode_charclass_class this_class;
4613 data_fake.flags = 0;
4615 data_fake.whilem_c = data->whilem_c;
4616 data_fake.last_closep = data->last_closep;
4619 data_fake.last_closep = &fake;
4620 data_fake.pos_delta = delta;
4621 if (flags & SCF_DO_STCLASS) {
4622 cl_init(pRExC_state, &this_class);
4623 data_fake.start_class = &this_class;
4624 f = SCF_DO_STCLASS_AND;
4626 if (flags & SCF_WHILEM_VISITED_POS)
4627 f |= SCF_WHILEM_VISITED_POS;
4629 if (trie->jump[word]) {
4631 nextbranch = trie_node + trie->jump[0];
4632 scan= trie_node + trie->jump[word];
4633 /* We go from the jump point to the branch that follows
4634 it. Note this means we need the vestigal unused branches
4635 even though they arent otherwise used.
4637 minnext = study_chunk(pRExC_state, &scan, minlenp,
4638 &deltanext, (regnode *)nextbranch, &data_fake,
4639 stopparen, recursed, NULL, f,depth+1);
4641 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
4642 nextbranch= regnext((regnode*)nextbranch);
4644 if (min1 > (I32)(minnext + trie->minlen))
4645 min1 = minnext + trie->minlen;
4646 if (deltanext == I32_MAX) {
4647 is_inf = is_inf_internal = 1;
4649 } else if (max1 < (I32)(minnext + deltanext + trie->maxlen))
4650 max1 = minnext + deltanext + trie->maxlen;
4652 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4654 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4655 if ( stopmin > min + min1)
4656 stopmin = min + min1;
4657 flags &= ~SCF_DO_SUBSTR;
4659 data->flags |= SCF_SEEN_ACCEPT;
4662 if (data_fake.flags & SF_HAS_EVAL)
4663 data->flags |= SF_HAS_EVAL;
4664 data->whilem_c = data_fake.whilem_c;
4666 if (flags & SCF_DO_STCLASS)
4667 cl_or(pRExC_state, &accum, &this_class);
4670 if (flags & SCF_DO_SUBSTR) {
4671 data->pos_min += min1;
4672 data->pos_delta += max1 - min1;
4673 if (max1 != min1 || is_inf)
4674 data->longest = &(data->longest_float);
4677 delta += max1 - min1;
4678 if (flags & SCF_DO_STCLASS_OR) {
4679 cl_or(pRExC_state, data->start_class, &accum);
4681 cl_and(data->start_class, and_withp);
4682 flags &= ~SCF_DO_STCLASS;
4685 else if (flags & SCF_DO_STCLASS_AND) {
4687 cl_and(data->start_class, &accum);
4688 flags &= ~SCF_DO_STCLASS;
4691 /* Switch to OR mode: cache the old value of
4692 * data->start_class */
4694 StructCopy(data->start_class, and_withp,
4695 struct regnode_charclass_class);
4696 flags &= ~SCF_DO_STCLASS_AND;
4697 StructCopy(&accum, data->start_class,
4698 struct regnode_charclass_class);
4699 flags |= SCF_DO_STCLASS_OR;
4700 SET_SSC_EOS(data->start_class);
4707 else if (PL_regkind[OP(scan)] == TRIE) {
4708 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4711 min += trie->minlen;
4712 delta += (trie->maxlen - trie->minlen);
4713 flags &= ~SCF_DO_STCLASS; /* xxx */
4714 if (flags & SCF_DO_SUBSTR) {
4715 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4716 data->pos_min += trie->minlen;
4717 data->pos_delta += (trie->maxlen - trie->minlen);
4718 if (trie->maxlen != trie->minlen)
4719 data->longest = &(data->longest_float);
4721 if (trie->jump) /* no more substrings -- for now /grr*/
4722 flags &= ~SCF_DO_SUBSTR;
4724 #endif /* old or new */
4725 #endif /* TRIE_STUDY_OPT */
4727 /* Else: zero-length, ignore. */
4728 scan = regnext(scan);
4733 stopparen = frame->stop;
4734 frame = frame->prev;
4735 goto fake_study_recurse;
4740 DEBUG_STUDYDATA("pre-fin:",data,depth);
4743 *deltap = is_inf_internal ? I32_MAX : delta;
4744 if (flags & SCF_DO_SUBSTR && is_inf)
4745 data->pos_delta = I32_MAX - data->pos_min;
4746 if (is_par > (I32)U8_MAX)
4748 if (is_par && pars==1 && data) {
4749 data->flags |= SF_IN_PAR;
4750 data->flags &= ~SF_HAS_PAR;
4752 else if (pars && data) {
4753 data->flags |= SF_HAS_PAR;
4754 data->flags &= ~SF_IN_PAR;
4756 if (flags & SCF_DO_STCLASS_OR)
4757 cl_and(data->start_class, and_withp);
4758 if (flags & SCF_TRIE_RESTUDY)
4759 data->flags |= SCF_TRIE_RESTUDY;
4761 DEBUG_STUDYDATA("post-fin:",data,depth);
4763 return min < stopmin ? min : stopmin;
4767 S_add_data(RExC_state_t *pRExC_state, U32 n, const char *s)
4769 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
4771 PERL_ARGS_ASSERT_ADD_DATA;
4773 Renewc(RExC_rxi->data,
4774 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
4775 char, struct reg_data);
4777 Renew(RExC_rxi->data->what, count + n, U8);
4779 Newx(RExC_rxi->data->what, n, U8);
4780 RExC_rxi->data->count = count + n;
4781 Copy(s, RExC_rxi->data->what + count, n, U8);
4785 /*XXX: todo make this not included in a non debugging perl */
4786 #ifndef PERL_IN_XSUB_RE
4788 Perl_reginitcolors(pTHX)
4791 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
4793 char *t = savepv(s);
4797 t = strchr(t, '\t');
4803 PL_colors[i] = t = (char *)"";
4808 PL_colors[i++] = (char *)"";
4815 #ifdef TRIE_STUDY_OPT
4816 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
4819 (data.flags & SCF_TRIE_RESTUDY) \
4827 #define CHECK_RESTUDY_GOTO_butfirst
4831 * pregcomp - compile a regular expression into internal code
4833 * Decides which engine's compiler to call based on the hint currently in
4837 #ifndef PERL_IN_XSUB_RE
4839 /* return the currently in-scope regex engine (or the default if none) */
4841 regexp_engine const *
4842 Perl_current_re_engine(pTHX)
4846 if (IN_PERL_COMPILETIME) {
4847 HV * const table = GvHV(PL_hintgv);
4851 return &PL_core_reg_engine;
4852 ptr = hv_fetchs(table, "regcomp", FALSE);
4853 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
4854 return &PL_core_reg_engine;
4855 return INT2PTR(regexp_engine*,SvIV(*ptr));
4859 if (!PL_curcop->cop_hints_hash)
4860 return &PL_core_reg_engine;
4861 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
4862 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
4863 return &PL_core_reg_engine;
4864 return INT2PTR(regexp_engine*,SvIV(ptr));
4870 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
4873 regexp_engine const *eng = current_re_engine();
4874 GET_RE_DEBUG_FLAGS_DECL;
4876 PERL_ARGS_ASSERT_PREGCOMP;
4878 /* Dispatch a request to compile a regexp to correct regexp engine. */
4880 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
4883 return CALLREGCOMP_ENG(eng, pattern, flags);
4887 /* public(ish) entry point for the perl core's own regex compiling code.
4888 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
4889 * pattern rather than a list of OPs, and uses the internal engine rather
4890 * than the current one */
4893 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
4895 SV *pat = pattern; /* defeat constness! */
4896 PERL_ARGS_ASSERT_RE_COMPILE;
4897 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
4898 #ifdef PERL_IN_XSUB_RE
4901 &PL_core_reg_engine,
4903 NULL, NULL, rx_flags, 0);
4907 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
4908 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
4909 * point to the realloced string and length.
4911 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
4915 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
4916 char **pat_p, STRLEN *plen_p, int num_code_blocks)
4918 U8 *const src = (U8*)*pat_p;
4921 STRLEN s = 0, d = 0;
4923 GET_RE_DEBUG_FLAGS_DECL;
4925 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
4926 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
4928 Newx(dst, *plen_p * 2 + 1, U8);
4930 while (s < *plen_p) {
4931 const UV uv = NATIVE_TO_ASCII(src[s]);
4932 if (UNI_IS_INVARIANT(uv))
4933 dst[d] = (U8)UTF_TO_NATIVE(uv);
4935 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
4936 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
4938 if (n < num_code_blocks) {
4939 if (!do_end && pRExC_state->code_blocks[n].start == s) {
4940 pRExC_state->code_blocks[n].start = d;
4941 assert(dst[d] == '(');
4944 else if (do_end && pRExC_state->code_blocks[n].end == s) {
4945 pRExC_state->code_blocks[n].end = d;
4946 assert(dst[d] == ')');
4956 *pat_p = (char*) dst;
4958 RExC_orig_utf8 = RExC_utf8 = 1;
4963 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
4964 * while recording any code block indices, and handling overloading,
4965 * nested qr// objects etc. If pat is null, it will allocate a new
4966 * string, or just return the first arg, if there's only one.
4968 * Returns the malloced/updated pat.
4969 * patternp and pat_count is the array of SVs to be concatted;
4970 * oplist is the optional list of ops that generated the SVs;
4971 * recompile_p is a pointer to a boolean that will be set if
4972 * the regex will need to be recompiled.
4973 * delim, if non-null is an SV that will be inserted between each element
4977 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
4978 SV *pat, SV ** const patternp, int pat_count,
4979 OP *oplist, bool *recompile_p, SV *delim)
4983 bool use_delim = FALSE;
4984 bool alloced = FALSE;
4986 /* if we know we have at least two args, create an empty string,
4987 * then concatenate args to that. For no args, return an empty string */
4988 if (!pat && pat_count != 1) {
4989 pat = newSVpvn("", 0);
4994 for (svp = patternp; svp < patternp + pat_count; svp++) {
4997 STRLEN orig_patlen = 0;
4999 SV *msv = use_delim ? delim : *svp;
5001 /* if we've got a delimiter, we go round the loop twice for each
5002 * svp slot (except the last), using the delimiter the second
5011 if (SvTYPE(msv) == SVt_PVAV) {
5012 /* we've encountered an interpolated array within
5013 * the pattern, e.g. /...@a..../. Expand the list of elements,
5014 * then recursively append elements.
5015 * The code in this block is based on S_pushav() */
5017 AV *const av = (AV*)msv;
5018 const I32 maxarg = AvFILL(av) + 1;
5022 assert(oplist->op_type == OP_PADAV
5023 || oplist->op_type == OP_RV2AV);
5024 oplist = oplist->op_sibling;;
5027 if (SvRMAGICAL(av)) {
5030 Newx(array, maxarg, SV*);
5032 for (i=0; i < (U32)maxarg; i++) {
5033 SV ** const svp = av_fetch(av, i, FALSE);
5034 array[i] = svp ? *svp : &PL_sv_undef;
5038 array = AvARRAY(av);
5040 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5041 array, maxarg, NULL, recompile_p,
5043 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5049 /* we make the assumption here that each op in the list of
5050 * op_siblings maps to one SV pushed onto the stack,
5051 * except for code blocks, with have both an OP_NULL and
5053 * This allows us to match up the list of SVs against the
5054 * list of OPs to find the next code block.
5056 * Note that PUSHMARK PADSV PADSV ..
5058 * PADRANGE PADSV PADSV ..
5059 * so the alignment still works. */
5062 if (oplist->op_type == OP_NULL
5063 && (oplist->op_flags & OPf_SPECIAL))
5065 assert(n < pRExC_state->num_code_blocks);
5066 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5067 pRExC_state->code_blocks[n].block = oplist;
5068 pRExC_state->code_blocks[n].src_regex = NULL;
5071 oplist = oplist->op_sibling; /* skip CONST */
5074 oplist = oplist->op_sibling;;
5077 /* apply magic and QR overloading to arg */
5080 if (SvROK(msv) && SvAMAGIC(msv)) {
5081 SV *sv = AMG_CALLunary(msv, regexp_amg);
5085 if (SvTYPE(sv) != SVt_REGEXP)
5086 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5091 /* try concatenation overload ... */
5092 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5093 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5096 /* overloading involved: all bets are off over literal
5097 * code. Pretend we haven't seen it */
5098 pRExC_state->num_code_blocks -= n;
5102 /* ... or failing that, try "" overload */
5103 while (SvAMAGIC(msv)
5104 && (sv = AMG_CALLunary(msv, string_amg))
5108 && SvRV(msv) == SvRV(sv))
5113 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5117 /* this is a partially unrolled
5118 * sv_catsv_nomg(pat, msv);
5119 * that allows us to adjust code block indices if
5122 char *dst = SvPV_force_nomg(pat, dlen);
5124 if (SvUTF8(msv) && !SvUTF8(pat)) {
5125 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5126 sv_setpvn(pat, dst, dlen);
5129 sv_catsv_nomg(pat, msv);
5136 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5139 /* extract any code blocks within any embedded qr//'s */
5140 if (rx && SvTYPE(rx) == SVt_REGEXP
5141 && RX_ENGINE((REGEXP*)rx)->op_comp)
5144 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5145 if (ri->num_code_blocks) {
5147 /* the presence of an embedded qr// with code means
5148 * we should always recompile: the text of the
5149 * qr// may not have changed, but it may be a
5150 * different closure than last time */
5152 Renew(pRExC_state->code_blocks,
5153 pRExC_state->num_code_blocks + ri->num_code_blocks,
5154 struct reg_code_block);
5155 pRExC_state->num_code_blocks += ri->num_code_blocks;
5157 for (i=0; i < ri->num_code_blocks; i++) {
5158 struct reg_code_block *src, *dst;
5159 STRLEN offset = orig_patlen
5160 + ReANY((REGEXP *)rx)->pre_prefix;
5161 assert(n < pRExC_state->num_code_blocks);
5162 src = &ri->code_blocks[i];
5163 dst = &pRExC_state->code_blocks[n];
5164 dst->start = src->start + offset;
5165 dst->end = src->end + offset;
5166 dst->block = src->block;
5167 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5176 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5185 /* see if there are any run-time code blocks in the pattern.
5186 * False positives are allowed */
5189 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5190 char *pat, STRLEN plen)
5195 for (s = 0; s < plen; s++) {
5196 if (n < pRExC_state->num_code_blocks
5197 && s == pRExC_state->code_blocks[n].start)
5199 s = pRExC_state->code_blocks[n].end;
5203 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5205 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5207 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5214 /* Handle run-time code blocks. We will already have compiled any direct
5215 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5216 * copy of it, but with any literal code blocks blanked out and
5217 * appropriate chars escaped; then feed it into
5219 * eval "qr'modified_pattern'"
5223 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5227 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5229 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5230 * and merge them with any code blocks of the original regexp.
5232 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5233 * instead, just save the qr and return FALSE; this tells our caller that
5234 * the original pattern needs upgrading to utf8.
5238 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5239 char *pat, STRLEN plen)
5243 GET_RE_DEBUG_FLAGS_DECL;
5245 if (pRExC_state->runtime_code_qr) {
5246 /* this is the second time we've been called; this should
5247 * only happen if the main pattern got upgraded to utf8
5248 * during compilation; re-use the qr we compiled first time
5249 * round (which should be utf8 too)
5251 qr = pRExC_state->runtime_code_qr;
5252 pRExC_state->runtime_code_qr = NULL;
5253 assert(RExC_utf8 && SvUTF8(qr));
5259 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5263 /* determine how many extra chars we need for ' and \ escaping */
5264 for (s = 0; s < plen; s++) {
5265 if (pat[s] == '\'' || pat[s] == '\\')
5269 Newx(newpat, newlen, char);
5271 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5273 for (s = 0; s < plen; s++) {
5274 if (n < pRExC_state->num_code_blocks
5275 && s == pRExC_state->code_blocks[n].start)
5277 /* blank out literal code block */
5278 assert(pat[s] == '(');
5279 while (s <= pRExC_state->code_blocks[n].end) {
5287 if (pat[s] == '\'' || pat[s] == '\\')
5292 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5296 PerlIO_printf(Perl_debug_log,
5297 "%sre-parsing pattern for runtime code:%s %s\n",
5298 PL_colors[4],PL_colors[5],newpat);
5301 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5307 PUSHSTACKi(PERLSI_REQUIRE);
5308 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5309 * parsing qr''; normally only q'' does this. It also alters
5311 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5312 SvREFCNT_dec_NN(sv);
5317 SV * const errsv = ERRSV;
5318 if (SvTRUE_NN(errsv))
5320 Safefree(pRExC_state->code_blocks);
5321 /* use croak_sv ? */
5322 Perl_croak_nocontext("%s", SvPV_nolen_const(errsv));
5325 assert(SvROK(qr_ref));
5327 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5328 /* the leaving below frees the tmp qr_ref.
5329 * Give qr a life of its own */
5337 if (!RExC_utf8 && SvUTF8(qr)) {
5338 /* first time through; the pattern got upgraded; save the
5339 * qr for the next time through */
5340 assert(!pRExC_state->runtime_code_qr);
5341 pRExC_state->runtime_code_qr = qr;
5346 /* extract any code blocks within the returned qr// */
5349 /* merge the main (r1) and run-time (r2) code blocks into one */
5351 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5352 struct reg_code_block *new_block, *dst;
5353 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5356 if (!r2->num_code_blocks) /* we guessed wrong */
5358 SvREFCNT_dec_NN(qr);
5363 r1->num_code_blocks + r2->num_code_blocks,
5364 struct reg_code_block);
5367 while ( i1 < r1->num_code_blocks
5368 || i2 < r2->num_code_blocks)
5370 struct reg_code_block *src;
5373 if (i1 == r1->num_code_blocks) {
5374 src = &r2->code_blocks[i2++];
5377 else if (i2 == r2->num_code_blocks)
5378 src = &r1->code_blocks[i1++];
5379 else if ( r1->code_blocks[i1].start
5380 < r2->code_blocks[i2].start)
5382 src = &r1->code_blocks[i1++];
5383 assert(src->end < r2->code_blocks[i2].start);
5386 assert( r1->code_blocks[i1].start
5387 > r2->code_blocks[i2].start);
5388 src = &r2->code_blocks[i2++];
5390 assert(src->end < r1->code_blocks[i1].start);
5393 assert(pat[src->start] == '(');
5394 assert(pat[src->end] == ')');
5395 dst->start = src->start;
5396 dst->end = src->end;
5397 dst->block = src->block;
5398 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5402 r1->num_code_blocks += r2->num_code_blocks;
5403 Safefree(r1->code_blocks);
5404 r1->code_blocks = new_block;
5407 SvREFCNT_dec_NN(qr);
5413 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest, SV** rx_utf8, SV** rx_substr, I32* rx_end_shift, I32 lookbehind, I32 offset, I32 *minlen, STRLEN longest_length, bool eol, bool meol)
5415 /* This is the common code for setting up the floating and fixed length
5416 * string data extracted from Perl_re_op_compile() below. Returns a boolean
5417 * as to whether succeeded or not */
5421 if (! (longest_length
5422 || (eol /* Can't have SEOL and MULTI */
5423 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5425 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5426 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5431 /* copy the information about the longest from the reg_scan_data
5432 over to the program. */
5433 if (SvUTF8(sv_longest)) {
5434 *rx_utf8 = sv_longest;
5437 *rx_substr = sv_longest;
5440 /* end_shift is how many chars that must be matched that
5441 follow this item. We calculate it ahead of time as once the
5442 lookbehind offset is added in we lose the ability to correctly
5444 ml = minlen ? *(minlen) : (I32)longest_length;
5445 *rx_end_shift = ml - offset
5446 - longest_length + (SvTAIL(sv_longest) != 0)
5449 t = (eol/* Can't have SEOL and MULTI */
5450 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5451 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5457 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5458 * regular expression into internal code.
5459 * The pattern may be passed either as:
5460 * a list of SVs (patternp plus pat_count)
5461 * a list of OPs (expr)
5462 * If both are passed, the SV list is used, but the OP list indicates
5463 * which SVs are actually pre-compiled code blocks
5465 * The SVs in the list have magic and qr overloading applied to them (and
5466 * the list may be modified in-place with replacement SVs in the latter
5469 * If the pattern hasn't changed from old_re, then old_re will be
5472 * eng is the current engine. If that engine has an op_comp method, then
5473 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5474 * do the initial concatenation of arguments and pass on to the external
5477 * If is_bare_re is not null, set it to a boolean indicating whether the
5478 * arg list reduced (after overloading) to a single bare regex which has
5479 * been returned (i.e. /$qr/).
5481 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5483 * pm_flags contains the PMf_* flags, typically based on those from the
5484 * pm_flags field of the related PMOP. Currently we're only interested in
5485 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5487 * We can't allocate space until we know how big the compiled form will be,
5488 * but we can't compile it (and thus know how big it is) until we've got a
5489 * place to put the code. So we cheat: we compile it twice, once with code
5490 * generation turned off and size counting turned on, and once "for real".
5491 * This also means that we don't allocate space until we are sure that the
5492 * thing really will compile successfully, and we never have to move the
5493 * code and thus invalidate pointers into it. (Note that it has to be in
5494 * one piece because free() must be able to free it all.) [NB: not true in perl]
5496 * Beware that the optimization-preparation code in here knows about some
5497 * of the structure of the compiled regexp. [I'll say.]
5501 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5502 OP *expr, const regexp_engine* eng, REGEXP *old_re,
5503 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5508 regexp_internal *ri;
5516 SV *code_blocksv = NULL;
5517 SV** new_patternp = patternp;
5519 /* these are all flags - maybe they should be turned
5520 * into a single int with different bit masks */
5521 I32 sawlookahead = 0;
5526 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5528 bool runtime_code = 0;
5530 RExC_state_t RExC_state;
5531 RExC_state_t * const pRExC_state = &RExC_state;
5532 #ifdef TRIE_STUDY_OPT
5534 RExC_state_t copyRExC_state;
5536 GET_RE_DEBUG_FLAGS_DECL;
5538 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5540 DEBUG_r(if (!PL_colorset) reginitcolors());
5542 #ifndef PERL_IN_XSUB_RE
5543 /* Initialize these here instead of as-needed, as is quick and avoids
5544 * having to test them each time otherwise */
5545 if (! PL_AboveLatin1) {
5546 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5547 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5548 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5550 PL_L1Posix_ptrs[_CC_ALPHANUMERIC]
5551 = _new_invlist_C_array(L1PosixAlnum_invlist);
5552 PL_Posix_ptrs[_CC_ALPHANUMERIC]
5553 = _new_invlist_C_array(PosixAlnum_invlist);
5555 PL_L1Posix_ptrs[_CC_ALPHA]
5556 = _new_invlist_C_array(L1PosixAlpha_invlist);
5557 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(PosixAlpha_invlist);
5559 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(PosixBlank_invlist);
5560 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(XPosixBlank_invlist);
5562 /* Cased is the same as Alpha in the ASCII range */
5563 PL_L1Posix_ptrs[_CC_CASED] = _new_invlist_C_array(L1Cased_invlist);
5564 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(PosixAlpha_invlist);
5566 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(PosixCntrl_invlist);
5567 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(XPosixCntrl_invlist);
5569 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5570 PL_L1Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5572 PL_L1Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(L1PosixGraph_invlist);
5573 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(PosixGraph_invlist);
5575 PL_L1Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(L1PosixLower_invlist);
5576 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(PosixLower_invlist);
5578 PL_L1Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(L1PosixPrint_invlist);
5579 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(PosixPrint_invlist);
5581 PL_L1Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(L1PosixPunct_invlist);
5582 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(PosixPunct_invlist);
5584 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(PerlSpace_invlist);
5585 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(XPerlSpace_invlist);
5586 PL_Posix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(PosixSpace_invlist);
5587 PL_XPosix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(XPosixSpace_invlist);
5589 PL_L1Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(L1PosixUpper_invlist);
5590 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(PosixUpper_invlist);
5592 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(VertSpace_invlist);
5594 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(PosixWord_invlist);
5595 PL_L1Posix_ptrs[_CC_WORDCHAR]
5596 = _new_invlist_C_array(L1PosixWord_invlist);
5598 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(PosixXDigit_invlist);
5599 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(XPosixXDigit_invlist);
5601 PL_HasMultiCharFold = _new_invlist_C_array(_Perl_Multi_Char_Folds_invlist);
5605 pRExC_state->code_blocks = NULL;
5606 pRExC_state->num_code_blocks = 0;
5609 *is_bare_re = FALSE;
5611 if (expr && (expr->op_type == OP_LIST ||
5612 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5613 /* allocate code_blocks if needed */
5617 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
5618 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5619 ncode++; /* count of DO blocks */
5621 pRExC_state->num_code_blocks = ncode;
5622 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5627 /* compile-time pattern with just OP_CONSTs and DO blocks */
5632 /* find how many CONSTs there are */
5635 if (expr->op_type == OP_CONST)
5638 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5639 if (o->op_type == OP_CONST)
5643 /* fake up an SV array */
5645 assert(!new_patternp);
5646 Newx(new_patternp, n, SV*);
5647 SAVEFREEPV(new_patternp);
5651 if (expr->op_type == OP_CONST)
5652 new_patternp[n] = cSVOPx_sv(expr);
5654 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5655 if (o->op_type == OP_CONST)
5656 new_patternp[n++] = cSVOPo_sv;
5661 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5662 "Assembling pattern from %d elements%s\n", pat_count,
5663 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5665 /* set expr to the first arg op */
5667 if (pRExC_state->num_code_blocks
5668 && expr->op_type != OP_CONST)
5670 expr = cLISTOPx(expr)->op_first;
5671 assert( expr->op_type == OP_PUSHMARK
5672 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
5673 || expr->op_type == OP_PADRANGE);
5674 expr = expr->op_sibling;
5677 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
5678 expr, &recompile, NULL);
5680 /* handle bare (possibly after overloading) regex: foo =~ $re */
5685 if (SvTYPE(re) == SVt_REGEXP) {
5689 Safefree(pRExC_state->code_blocks);
5690 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5691 "Precompiled pattern%s\n",
5692 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5698 exp = SvPV_nomg(pat, plen);
5700 if (!eng->op_comp) {
5701 if ((SvUTF8(pat) && IN_BYTES)
5702 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5704 /* make a temporary copy; either to convert to bytes,
5705 * or to avoid repeating get-magic / overloaded stringify */
5706 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5707 (IN_BYTES ? 0 : SvUTF8(pat)));
5709 Safefree(pRExC_state->code_blocks);
5710 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5713 /* ignore the utf8ness if the pattern is 0 length */
5714 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5715 RExC_uni_semantics = 0;
5716 RExC_contains_locale = 0;
5717 pRExC_state->runtime_code_qr = NULL;
5720 SV *dsv= sv_newmortal();
5721 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
5722 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5723 PL_colors[4],PL_colors[5],s);
5727 /* we jump here if we upgrade the pattern to utf8 and have to
5730 if ((pm_flags & PMf_USE_RE_EVAL)
5731 /* this second condition covers the non-regex literal case,
5732 * i.e. $foo =~ '(?{})'. */
5733 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
5735 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
5737 /* return old regex if pattern hasn't changed */
5738 /* XXX: note in the below we have to check the flags as well as the pattern.
5740 * Things get a touch tricky as we have to compare the utf8 flag independently
5741 * from the compile flags.
5746 && !!RX_UTF8(old_re) == !!RExC_utf8
5747 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
5748 && RX_PRECOMP(old_re)
5749 && RX_PRELEN(old_re) == plen
5750 && memEQ(RX_PRECOMP(old_re), exp, plen)
5751 && !runtime_code /* with runtime code, always recompile */ )
5753 Safefree(pRExC_state->code_blocks);
5757 rx_flags = orig_rx_flags;
5759 if (initial_charset == REGEX_LOCALE_CHARSET) {
5760 RExC_contains_locale = 1;
5762 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5764 /* Set to use unicode semantics if the pattern is in utf8 and has the
5765 * 'depends' charset specified, as it means unicode when utf8 */
5766 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5770 RExC_flags = rx_flags;
5771 RExC_pm_flags = pm_flags;
5774 if (TAINTING_get && TAINT_get)
5775 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5777 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5778 /* whoops, we have a non-utf8 pattern, whilst run-time code
5779 * got compiled as utf8. Try again with a utf8 pattern */
5780 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5781 pRExC_state->num_code_blocks);
5782 goto redo_first_pass;
5785 assert(!pRExC_state->runtime_code_qr);
5790 RExC_in_lookbehind = 0;
5791 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5793 RExC_override_recoding = 0;
5794 RExC_in_multi_char_class = 0;
5796 /* First pass: determine size, legality. */
5799 RExC_end = exp + plen;
5804 RExC_emit = &RExC_emit_dummy;
5805 RExC_whilem_seen = 0;
5806 RExC_open_parens = NULL;
5807 RExC_close_parens = NULL;
5809 RExC_paren_names = NULL;
5811 RExC_paren_name_list = NULL;
5813 RExC_recurse = NULL;
5814 RExC_recurse_count = 0;
5815 pRExC_state->code_index = 0;
5817 #if 0 /* REGC() is (currently) a NOP at the first pass.
5818 * Clever compilers notice this and complain. --jhi */
5819 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5822 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5824 RExC_lastparse=NULL;
5826 /* reg may croak on us, not giving us a chance to free
5827 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
5828 need it to survive as long as the regexp (qr/(?{})/).
5829 We must check that code_blocksv is not already set, because we may
5830 have jumped back to restart the sizing pass. */
5831 if (pRExC_state->code_blocks && !code_blocksv) {
5832 code_blocksv = newSV_type(SVt_PV);
5833 SAVEFREESV(code_blocksv);
5834 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
5835 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
5837 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5838 /* It's possible to write a regexp in ascii that represents Unicode
5839 codepoints outside of the byte range, such as via \x{100}. If we
5840 detect such a sequence we have to convert the entire pattern to utf8
5841 and then recompile, as our sizing calculation will have been based
5842 on 1 byte == 1 character, but we will need to use utf8 to encode
5843 at least some part of the pattern, and therefore must convert the whole
5846 if (flags & RESTART_UTF8) {
5847 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5848 pRExC_state->num_code_blocks);
5849 goto redo_first_pass;
5851 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
5854 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
5857 PerlIO_printf(Perl_debug_log,
5858 "Required size %"IVdf" nodes\n"
5859 "Starting second pass (creation)\n",
5862 RExC_lastparse=NULL;
5865 /* The first pass could have found things that force Unicode semantics */
5866 if ((RExC_utf8 || RExC_uni_semantics)
5867 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5869 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5872 /* Small enough for pointer-storage convention?
5873 If extralen==0, this means that we will not need long jumps. */
5874 if (RExC_size >= 0x10000L && RExC_extralen)
5875 RExC_size += RExC_extralen;
5878 if (RExC_whilem_seen > 15)
5879 RExC_whilem_seen = 15;
5881 /* Allocate space and zero-initialize. Note, the two step process
5882 of zeroing when in debug mode, thus anything assigned has to
5883 happen after that */
5884 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5886 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5887 char, regexp_internal);
5888 if ( r == NULL || ri == NULL )
5889 FAIL("Regexp out of space");
5891 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5892 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5894 /* bulk initialize base fields with 0. */
5895 Zero(ri, sizeof(regexp_internal), char);
5898 /* non-zero initialization begins here */
5901 r->extflags = rx_flags;
5902 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
5904 if (pm_flags & PMf_IS_QR) {
5905 ri->code_blocks = pRExC_state->code_blocks;
5906 ri->num_code_blocks = pRExC_state->num_code_blocks;
5911 for (n = 0; n < pRExC_state->num_code_blocks; n++)
5912 if (pRExC_state->code_blocks[n].src_regex)
5913 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
5914 SAVEFREEPV(pRExC_state->code_blocks);
5918 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5919 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5921 /* The caret is output if there are any defaults: if not all the STD
5922 * flags are set, or if no character set specifier is needed */
5924 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5926 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5927 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5928 >> RXf_PMf_STD_PMMOD_SHIFT);
5929 const char *fptr = STD_PAT_MODS; /*"msix"*/
5931 /* Allocate for the worst case, which is all the std flags are turned
5932 * on. If more precision is desired, we could do a population count of
5933 * the flags set. This could be done with a small lookup table, or by
5934 * shifting, masking and adding, or even, when available, assembly
5935 * language for a machine-language population count.
5936 * We never output a minus, as all those are defaults, so are
5937 * covered by the caret */
5938 const STRLEN wraplen = plen + has_p + has_runon
5939 + has_default /* If needs a caret */
5941 /* If needs a character set specifier */
5942 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5943 + (sizeof(STD_PAT_MODS) - 1)
5944 + (sizeof("(?:)") - 1);
5946 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
5947 r->xpv_len_u.xpvlenu_pv = p;
5949 SvFLAGS(rx) |= SVf_UTF8;
5952 /* If a default, cover it using the caret */
5954 *p++= DEFAULT_PAT_MOD;
5958 const char* const name = get_regex_charset_name(r->extflags, &len);
5959 Copy(name, p, len, char);
5963 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5966 while((ch = *fptr++)) {
5974 Copy(RExC_precomp, p, plen, char);
5975 assert ((RX_WRAPPED(rx) - p) < 16);
5976 r->pre_prefix = p - RX_WRAPPED(rx);
5982 SvCUR_set(rx, p - RX_WRAPPED(rx));
5986 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5988 if (RExC_seen & REG_SEEN_RECURSE) {
5989 Newxz(RExC_open_parens, RExC_npar,regnode *);
5990 SAVEFREEPV(RExC_open_parens);
5991 Newxz(RExC_close_parens,RExC_npar,regnode *);
5992 SAVEFREEPV(RExC_close_parens);
5995 /* Useful during FAIL. */
5996 #ifdef RE_TRACK_PATTERN_OFFSETS
5997 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5998 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
5999 "%s %"UVuf" bytes for offset annotations.\n",
6000 ri->u.offsets ? "Got" : "Couldn't get",
6001 (UV)((2*RExC_size+1) * sizeof(U32))));
6003 SetProgLen(ri,RExC_size);
6008 /* Second pass: emit code. */
6009 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6010 RExC_pm_flags = pm_flags;
6012 RExC_end = exp + plen;
6015 RExC_emit_start = ri->program;
6016 RExC_emit = ri->program;
6017 RExC_emit_bound = ri->program + RExC_size + 1;
6018 pRExC_state->code_index = 0;
6020 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
6021 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6023 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6025 /* XXXX To minimize changes to RE engine we always allocate
6026 3-units-long substrs field. */
6027 Newx(r->substrs, 1, struct reg_substr_data);
6028 if (RExC_recurse_count) {
6029 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6030 SAVEFREEPV(RExC_recurse);
6034 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
6035 Zero(r->substrs, 1, struct reg_substr_data);
6037 #ifdef TRIE_STUDY_OPT
6039 StructCopy(&zero_scan_data, &data, scan_data_t);
6040 copyRExC_state = RExC_state;
6043 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6045 RExC_state = copyRExC_state;
6046 if (seen & REG_TOP_LEVEL_BRANCHES)
6047 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
6049 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
6050 StructCopy(&zero_scan_data, &data, scan_data_t);
6053 StructCopy(&zero_scan_data, &data, scan_data_t);
6056 /* Dig out information for optimizations. */
6057 r->extflags = RExC_flags; /* was pm_op */
6058 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6061 SvUTF8_on(rx); /* Unicode in it? */
6062 ri->regstclass = NULL;
6063 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6064 r->intflags |= PREGf_NAUGHTY;
6065 scan = ri->program + 1; /* First BRANCH. */
6067 /* testing for BRANCH here tells us whether there is "must appear"
6068 data in the pattern. If there is then we can use it for optimisations */
6069 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
6071 STRLEN longest_float_length, longest_fixed_length;
6072 struct regnode_charclass_class ch_class; /* pointed to by data */
6074 I32 last_close = 0; /* pointed to by data */
6075 regnode *first= scan;
6076 regnode *first_next= regnext(first);
6078 * Skip introductions and multiplicators >= 1
6079 * so that we can extract the 'meat' of the pattern that must
6080 * match in the large if() sequence following.
6081 * NOTE that EXACT is NOT covered here, as it is normally
6082 * picked up by the optimiser separately.
6084 * This is unfortunate as the optimiser isnt handling lookahead
6085 * properly currently.
6088 while ((OP(first) == OPEN && (sawopen = 1)) ||
6089 /* An OR of *one* alternative - should not happen now. */
6090 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6091 /* for now we can't handle lookbehind IFMATCH*/
6092 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6093 (OP(first) == PLUS) ||
6094 (OP(first) == MINMOD) ||
6095 /* An {n,m} with n>0 */
6096 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6097 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6100 * the only op that could be a regnode is PLUS, all the rest
6101 * will be regnode_1 or regnode_2.
6103 * (yves doesn't think this is true)
6105 if (OP(first) == PLUS)
6108 if (OP(first) == MINMOD)
6110 first += regarglen[OP(first)];
6112 first = NEXTOPER(first);
6113 first_next= regnext(first);
6116 /* Starting-point info. */
6118 DEBUG_PEEP("first:",first,0);
6119 /* Ignore EXACT as we deal with it later. */
6120 if (PL_regkind[OP(first)] == EXACT) {
6121 if (OP(first) == EXACT)
6122 NOOP; /* Empty, get anchored substr later. */
6124 ri->regstclass = first;
6127 else if (PL_regkind[OP(first)] == TRIE &&
6128 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6131 /* this can happen only on restudy */
6132 if ( OP(first) == TRIE ) {
6133 struct regnode_1 *trieop = (struct regnode_1 *)
6134 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6135 StructCopy(first,trieop,struct regnode_1);
6136 trie_op=(regnode *)trieop;
6138 struct regnode_charclass *trieop = (struct regnode_charclass *)
6139 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6140 StructCopy(first,trieop,struct regnode_charclass);
6141 trie_op=(regnode *)trieop;
6144 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6145 ri->regstclass = trie_op;
6148 else if (REGNODE_SIMPLE(OP(first)))
6149 ri->regstclass = first;
6150 else if (PL_regkind[OP(first)] == BOUND ||
6151 PL_regkind[OP(first)] == NBOUND)
6152 ri->regstclass = first;
6153 else if (PL_regkind[OP(first)] == BOL) {
6154 r->extflags |= (OP(first) == MBOL
6156 : (OP(first) == SBOL
6159 first = NEXTOPER(first);
6162 else if (OP(first) == GPOS) {
6163 r->extflags |= RXf_ANCH_GPOS;
6164 first = NEXTOPER(first);
6167 else if ((!sawopen || !RExC_sawback) &&
6168 (OP(first) == STAR &&
6169 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6170 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6172 /* turn .* into ^.* with an implied $*=1 */
6174 (OP(NEXTOPER(first)) == REG_ANY)
6177 r->extflags |= type;
6178 r->intflags |= PREGf_IMPLICIT;
6179 first = NEXTOPER(first);
6182 if (sawplus && !sawminmod && !sawlookahead && (!sawopen || !RExC_sawback)
6183 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6184 /* x+ must match at the 1st pos of run of x's */
6185 r->intflags |= PREGf_SKIP;
6187 /* Scan is after the zeroth branch, first is atomic matcher. */
6188 #ifdef TRIE_STUDY_OPT
6191 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6192 (IV)(first - scan + 1))
6196 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6197 (IV)(first - scan + 1))
6203 * If there's something expensive in the r.e., find the
6204 * longest literal string that must appear and make it the
6205 * regmust. Resolve ties in favor of later strings, since
6206 * the regstart check works with the beginning of the r.e.
6207 * and avoiding duplication strengthens checking. Not a
6208 * strong reason, but sufficient in the absence of others.
6209 * [Now we resolve ties in favor of the earlier string if
6210 * it happens that c_offset_min has been invalidated, since the
6211 * earlier string may buy us something the later one won't.]
6214 data.longest_fixed = newSVpvs("");
6215 data.longest_float = newSVpvs("");
6216 data.last_found = newSVpvs("");
6217 data.longest = &(data.longest_fixed);
6218 ENTER_with_name("study_chunk");
6219 SAVEFREESV(data.longest_fixed);
6220 SAVEFREESV(data.longest_float);
6221 SAVEFREESV(data.last_found);
6223 if (!ri->regstclass) {
6224 cl_init(pRExC_state, &ch_class);
6225 data.start_class = &ch_class;
6226 stclass_flag = SCF_DO_STCLASS_AND;
6227 } else /* XXXX Check for BOUND? */
6229 data.last_closep = &last_close;
6231 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6232 &data, -1, NULL, NULL,
6233 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
6234 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6238 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6241 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6242 && data.last_start_min == 0 && data.last_end > 0
6243 && !RExC_seen_zerolen
6244 && !(RExC_seen & REG_SEEN_VERBARG)
6245 && !((RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6246 r->extflags |= RXf_CHECK_ALL;
6247 scan_commit(pRExC_state, &data,&minlen,0);
6249 longest_float_length = CHR_SVLEN(data.longest_float);
6251 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6252 && data.offset_fixed == data.offset_float_min
6253 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6254 && S_setup_longest (aTHX_ pRExC_state,
6258 &(r->float_end_shift),
6259 data.lookbehind_float,
6260 data.offset_float_min,
6262 longest_float_length,
6263 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6264 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6266 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6267 r->float_max_offset = data.offset_float_max;
6268 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6269 r->float_max_offset -= data.lookbehind_float;
6270 SvREFCNT_inc_simple_void_NN(data.longest_float);
6273 r->float_substr = r->float_utf8 = NULL;
6274 longest_float_length = 0;
6277 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6279 if (S_setup_longest (aTHX_ pRExC_state,
6281 &(r->anchored_utf8),
6282 &(r->anchored_substr),
6283 &(r->anchored_end_shift),
6284 data.lookbehind_fixed,
6287 longest_fixed_length,
6288 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6289 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6291 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6292 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6295 r->anchored_substr = r->anchored_utf8 = NULL;
6296 longest_fixed_length = 0;
6298 LEAVE_with_name("study_chunk");
6301 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6302 ri->regstclass = NULL;
6304 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6306 && ! TEST_SSC_EOS(data.start_class)
6307 && !cl_is_anything(data.start_class))
6309 const U32 n = add_data(pRExC_state, 1, "f");
6310 OP(data.start_class) = ANYOF_SYNTHETIC;
6312 Newx(RExC_rxi->data->data[n], 1,
6313 struct regnode_charclass_class);
6314 StructCopy(data.start_class,
6315 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6316 struct regnode_charclass_class);
6317 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6318 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6319 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6320 regprop(r, sv, (regnode*)data.start_class);
6321 PerlIO_printf(Perl_debug_log,
6322 "synthetic stclass \"%s\".\n",
6323 SvPVX_const(sv));});
6326 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6327 if (longest_fixed_length > longest_float_length) {
6328 r->check_end_shift = r->anchored_end_shift;
6329 r->check_substr = r->anchored_substr;
6330 r->check_utf8 = r->anchored_utf8;
6331 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6332 if (r->extflags & RXf_ANCH_SINGLE)
6333 r->extflags |= RXf_NOSCAN;
6336 r->check_end_shift = r->float_end_shift;
6337 r->check_substr = r->float_substr;
6338 r->check_utf8 = r->float_utf8;
6339 r->check_offset_min = r->float_min_offset;
6340 r->check_offset_max = r->float_max_offset;
6342 if ((r->check_substr || r->check_utf8) ) {
6343 r->extflags |= RXf_USE_INTUIT;
6344 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6345 r->extflags |= RXf_INTUIT_TAIL;
6347 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6348 if ( (STRLEN)minlen < longest_float_length )
6349 minlen= longest_float_length;
6350 if ( (STRLEN)minlen < longest_fixed_length )
6351 minlen= longest_fixed_length;
6355 /* Several toplevels. Best we can is to set minlen. */
6357 struct regnode_charclass_class ch_class;
6360 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6362 scan = ri->program + 1;
6363 cl_init(pRExC_state, &ch_class);
6364 data.start_class = &ch_class;
6365 data.last_closep = &last_close;
6368 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6369 &data, -1, NULL, NULL,
6370 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS
6371 |(restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6374 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6376 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6377 = r->float_substr = r->float_utf8 = NULL;
6379 if (! TEST_SSC_EOS(data.start_class)
6380 && !cl_is_anything(data.start_class))
6382 const U32 n = add_data(pRExC_state, 1, "f");
6383 OP(data.start_class) = ANYOF_SYNTHETIC;
6385 Newx(RExC_rxi->data->data[n], 1,
6386 struct regnode_charclass_class);
6387 StructCopy(data.start_class,
6388 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6389 struct regnode_charclass_class);
6390 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6391 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6392 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6393 regprop(r, sv, (regnode*)data.start_class);
6394 PerlIO_printf(Perl_debug_log,
6395 "synthetic stclass \"%s\".\n",
6396 SvPVX_const(sv));});
6400 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6401 the "real" pattern. */
6403 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6404 (IV)minlen, (IV)r->minlen);
6406 r->minlenret = minlen;
6407 if (r->minlen < minlen)
6410 if (RExC_seen & REG_SEEN_GPOS)
6411 r->extflags |= RXf_GPOS_SEEN;
6412 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6413 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the lookbehind */
6414 if (pRExC_state->num_code_blocks)
6415 r->extflags |= RXf_EVAL_SEEN;
6416 if (RExC_seen & REG_SEEN_CANY)
6417 r->extflags |= RXf_CANY_SEEN;
6418 if (RExC_seen & REG_SEEN_VERBARG)
6420 r->intflags |= PREGf_VERBARG_SEEN;
6421 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
6423 if (RExC_seen & REG_SEEN_CUTGROUP)
6424 r->intflags |= PREGf_CUTGROUP_SEEN;
6425 if (pm_flags & PMf_USE_RE_EVAL)
6426 r->intflags |= PREGf_USE_RE_EVAL;
6427 if (RExC_paren_names)
6428 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6430 RXp_PAREN_NAMES(r) = NULL;
6433 regnode *first = ri->program + 1;
6435 regnode *next = NEXTOPER(first);
6438 if (PL_regkind[fop] == NOTHING && nop == END)
6439 r->extflags |= RXf_NULL;
6440 else if (PL_regkind[fop] == BOL && nop == END)
6441 r->extflags |= RXf_START_ONLY;
6442 else if (fop == PLUS && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE && OP(regnext(first)) == END)
6443 r->extflags |= RXf_WHITE;
6444 else if ( r->extflags & RXf_SPLIT && fop == EXACT && STR_LEN(first) == 1 && *(STRING(first)) == ' ' && OP(regnext(first)) == END )
6445 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6449 if (RExC_paren_names) {
6450 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6451 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6454 ri->name_list_idx = 0;
6456 if (RExC_recurse_count) {
6457 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6458 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6459 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6462 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6463 /* assume we don't need to swap parens around before we match */
6466 PerlIO_printf(Perl_debug_log,"Final program:\n");
6469 #ifdef RE_TRACK_PATTERN_OFFSETS
6470 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6471 const U32 len = ri->u.offsets[0];
6473 GET_RE_DEBUG_FLAGS_DECL;
6474 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6475 for (i = 1; i <= len; i++) {
6476 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6477 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6478 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6480 PerlIO_printf(Perl_debug_log, "\n");
6485 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
6486 * by setting the regexp SV to readonly-only instead. If the
6487 * pattern's been recompiled, the USEDness should remain. */
6488 if (old_re && SvREADONLY(old_re))
6496 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6499 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6501 PERL_UNUSED_ARG(value);
6503 if (flags & RXapif_FETCH) {
6504 return reg_named_buff_fetch(rx, key, flags);
6505 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6506 Perl_croak_no_modify();
6508 } else if (flags & RXapif_EXISTS) {
6509 return reg_named_buff_exists(rx, key, flags)
6512 } else if (flags & RXapif_REGNAMES) {
6513 return reg_named_buff_all(rx, flags);
6514 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6515 return reg_named_buff_scalar(rx, flags);
6517 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6523 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6526 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6527 PERL_UNUSED_ARG(lastkey);
6529 if (flags & RXapif_FIRSTKEY)
6530 return reg_named_buff_firstkey(rx, flags);
6531 else if (flags & RXapif_NEXTKEY)
6532 return reg_named_buff_nextkey(rx, flags);
6534 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6540 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6543 AV *retarray = NULL;
6545 struct regexp *const rx = ReANY(r);
6547 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6549 if (flags & RXapif_ALL)
6552 if (rx && RXp_PAREN_NAMES(rx)) {
6553 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6556 SV* sv_dat=HeVAL(he_str);
6557 I32 *nums=(I32*)SvPVX(sv_dat);
6558 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6559 if ((I32)(rx->nparens) >= nums[i]
6560 && rx->offs[nums[i]].start != -1
6561 && rx->offs[nums[i]].end != -1)
6564 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6569 ret = newSVsv(&PL_sv_undef);
6572 av_push(retarray, ret);
6575 return newRV_noinc(MUTABLE_SV(retarray));
6582 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6585 struct regexp *const rx = ReANY(r);
6587 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6589 if (rx && RXp_PAREN_NAMES(rx)) {
6590 if (flags & RXapif_ALL) {
6591 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6593 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6595 SvREFCNT_dec_NN(sv);
6607 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6609 struct regexp *const rx = ReANY(r);
6611 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6613 if ( rx && RXp_PAREN_NAMES(rx) ) {
6614 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6616 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6623 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6625 struct regexp *const rx = ReANY(r);
6626 GET_RE_DEBUG_FLAGS_DECL;
6628 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6630 if (rx && RXp_PAREN_NAMES(rx)) {
6631 HV *hv = RXp_PAREN_NAMES(rx);
6633 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6636 SV* sv_dat = HeVAL(temphe);
6637 I32 *nums = (I32*)SvPVX(sv_dat);
6638 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6639 if ((I32)(rx->lastparen) >= nums[i] &&
6640 rx->offs[nums[i]].start != -1 &&
6641 rx->offs[nums[i]].end != -1)
6647 if (parno || flags & RXapif_ALL) {
6648 return newSVhek(HeKEY_hek(temphe));
6656 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6661 struct regexp *const rx = ReANY(r);
6663 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6665 if (rx && RXp_PAREN_NAMES(rx)) {
6666 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6667 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6668 } else if (flags & RXapif_ONE) {
6669 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6670 av = MUTABLE_AV(SvRV(ret));
6671 length = av_len(av);
6672 SvREFCNT_dec_NN(ret);
6673 return newSViv(length + 1);
6675 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6679 return &PL_sv_undef;
6683 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6685 struct regexp *const rx = ReANY(r);
6688 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6690 if (rx && RXp_PAREN_NAMES(rx)) {
6691 HV *hv= RXp_PAREN_NAMES(rx);
6693 (void)hv_iterinit(hv);
6694 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6697 SV* sv_dat = HeVAL(temphe);
6698 I32 *nums = (I32*)SvPVX(sv_dat);
6699 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6700 if ((I32)(rx->lastparen) >= nums[i] &&
6701 rx->offs[nums[i]].start != -1 &&
6702 rx->offs[nums[i]].end != -1)
6708 if (parno || flags & RXapif_ALL) {
6709 av_push(av, newSVhek(HeKEY_hek(temphe)));
6714 return newRV_noinc(MUTABLE_SV(av));
6718 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6721 struct regexp *const rx = ReANY(r);
6727 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6729 if ( n == RX_BUFF_IDX_CARET_PREMATCH
6730 || n == RX_BUFF_IDX_CARET_FULLMATCH
6731 || n == RX_BUFF_IDX_CARET_POSTMATCH
6734 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
6736 /* on something like
6739 * the KEEPCOPY is set on the PMOP rather than the regex */
6740 if (PL_curpm && r == PM_GETRE(PL_curpm))
6741 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
6750 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6751 /* no need to distinguish between them any more */
6752 n = RX_BUFF_IDX_FULLMATCH;
6754 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6755 && rx->offs[0].start != -1)
6757 /* $`, ${^PREMATCH} */
6758 i = rx->offs[0].start;
6762 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6763 && rx->offs[0].end != -1)
6765 /* $', ${^POSTMATCH} */
6766 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6767 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6770 if ( 0 <= n && n <= (I32)rx->nparens &&
6771 (s1 = rx->offs[n].start) != -1 &&
6772 (t1 = rx->offs[n].end) != -1)
6774 /* $&, ${^MATCH}, $1 ... */
6776 s = rx->subbeg + s1 - rx->suboffset;
6781 assert(s >= rx->subbeg);
6782 assert(rx->sublen >= (s - rx->subbeg) + i );
6784 #if NO_TAINT_SUPPORT
6785 sv_setpvn(sv, s, i);
6787 const int oldtainted = TAINT_get;
6789 sv_setpvn(sv, s, i);
6790 TAINT_set(oldtainted);
6792 if ( (rx->extflags & RXf_CANY_SEEN)
6793 ? (RXp_MATCH_UTF8(rx)
6794 && (!i || is_utf8_string((U8*)s, i)))
6795 : (RXp_MATCH_UTF8(rx)) )
6802 if (RXp_MATCH_TAINTED(rx)) {
6803 if (SvTYPE(sv) >= SVt_PVMG) {
6804 MAGIC* const mg = SvMAGIC(sv);
6807 SvMAGIC_set(sv, mg->mg_moremagic);
6809 if ((mgt = SvMAGIC(sv))) {
6810 mg->mg_moremagic = mgt;
6811 SvMAGIC_set(sv, mg);
6822 sv_setsv(sv,&PL_sv_undef);
6828 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6829 SV const * const value)
6831 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6833 PERL_UNUSED_ARG(rx);
6834 PERL_UNUSED_ARG(paren);
6835 PERL_UNUSED_ARG(value);
6838 Perl_croak_no_modify();
6842 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6845 struct regexp *const rx = ReANY(r);
6849 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6851 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
6852 || paren == RX_BUFF_IDX_CARET_FULLMATCH
6853 || paren == RX_BUFF_IDX_CARET_POSTMATCH
6856 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
6858 /* on something like
6861 * the KEEPCOPY is set on the PMOP rather than the regex */
6862 if (PL_curpm && r == PM_GETRE(PL_curpm))
6863 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
6869 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6871 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6872 case RX_BUFF_IDX_PREMATCH: /* $` */
6873 if (rx->offs[0].start != -1) {
6874 i = rx->offs[0].start;
6883 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6884 case RX_BUFF_IDX_POSTMATCH: /* $' */
6885 if (rx->offs[0].end != -1) {
6886 i = rx->sublen - rx->offs[0].end;
6888 s1 = rx->offs[0].end;
6895 default: /* $& / ${^MATCH}, $1, $2, ... */
6896 if (paren <= (I32)rx->nparens &&
6897 (s1 = rx->offs[paren].start) != -1 &&
6898 (t1 = rx->offs[paren].end) != -1)
6904 if (ckWARN(WARN_UNINITIALIZED))
6905 report_uninit((const SV *)sv);
6910 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6911 const char * const s = rx->subbeg - rx->suboffset + s1;
6916 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6923 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6925 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6926 PERL_UNUSED_ARG(rx);
6930 return newSVpvs("Regexp");
6933 /* Scans the name of a named buffer from the pattern.
6934 * If flags is REG_RSN_RETURN_NULL returns null.
6935 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6936 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6937 * to the parsed name as looked up in the RExC_paren_names hash.
6938 * If there is an error throws a vFAIL().. type exception.
6941 #define REG_RSN_RETURN_NULL 0
6942 #define REG_RSN_RETURN_NAME 1
6943 #define REG_RSN_RETURN_DATA 2
6946 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6948 char *name_start = RExC_parse;
6950 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6952 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6953 /* skip IDFIRST by using do...while */
6956 RExC_parse += UTF8SKIP(RExC_parse);
6957 } while (isWORDCHAR_utf8((U8*)RExC_parse));
6961 } while (isWORDCHAR(*RExC_parse));
6963 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6964 vFAIL("Group name must start with a non-digit word character");
6968 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6969 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6970 if ( flags == REG_RSN_RETURN_NAME)
6972 else if (flags==REG_RSN_RETURN_DATA) {
6975 if ( ! sv_name ) /* should not happen*/
6976 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6977 if (RExC_paren_names)
6978 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6980 sv_dat = HeVAL(he_str);
6982 vFAIL("Reference to nonexistent named group");
6986 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6987 (unsigned long) flags);
6989 assert(0); /* NOT REACHED */
6994 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6995 int rem=(int)(RExC_end - RExC_parse); \
7004 if (RExC_lastparse!=RExC_parse) \
7005 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
7008 iscut ? "..." : "<" \
7011 PerlIO_printf(Perl_debug_log,"%16s",""); \
7014 num = RExC_size + 1; \
7016 num=REG_NODE_NUM(RExC_emit); \
7017 if (RExC_lastnum!=num) \
7018 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7020 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7021 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7022 (int)((depth*2)), "", \
7026 RExC_lastparse=RExC_parse; \
7031 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7032 DEBUG_PARSE_MSG((funcname)); \
7033 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7035 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7036 DEBUG_PARSE_MSG((funcname)); \
7037 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7040 /* This section of code defines the inversion list object and its methods. The
7041 * interfaces are highly subject to change, so as much as possible is static to
7042 * this file. An inversion list is here implemented as a malloc'd C UV array
7043 * as an SVt_INVLIST scalar.
7045 * An inversion list for Unicode is an array of code points, sorted by ordinal
7046 * number. The zeroth element is the first code point in the list. The 1th
7047 * element is the first element beyond that not in the list. In other words,
7048 * the first range is
7049 * invlist[0]..(invlist[1]-1)
7050 * The other ranges follow. Thus every element whose index is divisible by two
7051 * marks the beginning of a range that is in the list, and every element not
7052 * divisible by two marks the beginning of a range not in the list. A single
7053 * element inversion list that contains the single code point N generally
7054 * consists of two elements
7057 * (The exception is when N is the highest representable value on the
7058 * machine, in which case the list containing just it would be a single
7059 * element, itself. By extension, if the last range in the list extends to
7060 * infinity, then the first element of that range will be in the inversion list
7061 * at a position that is divisible by two, and is the final element in the
7063 * Taking the complement (inverting) an inversion list is quite simple, if the
7064 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7065 * This implementation reserves an element at the beginning of each inversion
7066 * list to always contain 0; there is an additional flag in the header which
7067 * indicates if the list begins at the 0, or is offset to begin at the next
7070 * More about inversion lists can be found in "Unicode Demystified"
7071 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7072 * More will be coming when functionality is added later.
7074 * The inversion list data structure is currently implemented as an SV pointing
7075 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7076 * array of UV whose memory management is automatically handled by the existing
7077 * facilities for SV's.
7079 * Some of the methods should always be private to the implementation, and some
7080 * should eventually be made public */
7082 /* The header definitions are in F<inline_invlist.c> */
7084 PERL_STATIC_INLINE UV*
7085 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7087 /* Returns a pointer to the first element in the inversion list's array.
7088 * This is called upon initialization of an inversion list. Where the
7089 * array begins depends on whether the list has the code point U+0000 in it
7090 * or not. The other parameter tells it whether the code that follows this
7091 * call is about to put a 0 in the inversion list or not. The first
7092 * element is either the element reserved for 0, if TRUE, or the element
7093 * after it, if FALSE */
7095 bool* offset = get_invlist_offset_addr(invlist);
7096 UV* zero_addr = (UV *) SvPVX(invlist);
7098 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7101 assert(! _invlist_len(invlist));
7105 /* 1^1 = 0; 1^0 = 1 */
7106 *offset = 1 ^ will_have_0;
7107 return zero_addr + *offset;
7110 PERL_STATIC_INLINE UV*
7111 S_invlist_array(pTHX_ SV* const invlist)
7113 /* Returns the pointer to the inversion list's array. Every time the
7114 * length changes, this needs to be called in case malloc or realloc moved
7117 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7119 /* Must not be empty. If these fail, you probably didn't check for <len>
7120 * being non-zero before trying to get the array */
7121 assert(_invlist_len(invlist));
7123 /* The very first element always contains zero, The array begins either
7124 * there, or if the inversion list is offset, at the element after it.
7125 * The offset header field determines which; it contains 0 or 1 to indicate
7126 * how much additionally to add */
7127 assert(0 == *(SvPVX(invlist)));
7128 return ((UV *) SvPVX(invlist) + *get_invlist_offset_addr(invlist));
7131 PERL_STATIC_INLINE void
7132 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
7134 /* Sets the current number of elements stored in the inversion list.
7135 * Updates SvCUR correspondingly */
7137 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7139 assert(SvTYPE(invlist) == SVt_INVLIST);
7144 : TO_INTERNAL_SIZE(len + offset));
7145 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
7148 PERL_STATIC_INLINE IV*
7149 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7151 /* Return the address of the IV that is reserved to hold the cached index
7154 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7156 assert(SvTYPE(invlist) == SVt_INVLIST);
7158 return &(((XINVLIST*) SvANY(invlist))->prev_index);
7161 PERL_STATIC_INLINE IV
7162 S_invlist_previous_index(pTHX_ SV* const invlist)
7164 /* Returns cached index of previous search */
7166 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7168 return *get_invlist_previous_index_addr(invlist);
7171 PERL_STATIC_INLINE void
7172 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7174 /* Caches <index> for later retrieval */
7176 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7178 assert(index == 0 || index < (int) _invlist_len(invlist));
7180 *get_invlist_previous_index_addr(invlist) = index;
7183 PERL_STATIC_INLINE UV
7184 S_invlist_max(pTHX_ SV* const invlist)
7186 /* Returns the maximum number of elements storable in the inversion list's
7187 * array, without having to realloc() */
7189 PERL_ARGS_ASSERT_INVLIST_MAX;
7191 assert(SvTYPE(invlist) == SVt_INVLIST);
7193 /* Assumes worst case, in which the 0 element is not counted in the
7194 * inversion list, so subtracts 1 for that */
7195 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7196 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
7197 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
7200 #ifndef PERL_IN_XSUB_RE
7202 Perl__new_invlist(pTHX_ IV initial_size)
7205 /* Return a pointer to a newly constructed inversion list, with enough
7206 * space to store 'initial_size' elements. If that number is negative, a
7207 * system default is used instead */
7211 if (initial_size < 0) {
7215 /* Allocate the initial space */
7216 new_list = newSV_type(SVt_INVLIST);
7218 /* First 1 is in case the zero element isn't in the list; second 1 is for
7220 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
7221 invlist_set_len(new_list, 0, 0);
7223 /* Force iterinit() to be used to get iteration to work */
7224 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
7226 *get_invlist_previous_index_addr(new_list) = 0;
7233 S__new_invlist_C_array(pTHX_ const UV* const list)
7235 /* Return a pointer to a newly constructed inversion list, initialized to
7236 * point to <list>, which has to be in the exact correct inversion list
7237 * form, including internal fields. Thus this is a dangerous routine that
7238 * should not be used in the wrong hands. The passed in 'list' contains
7239 * several header fields at the beginning that are not part of the
7240 * inversion list body proper */
7242 const STRLEN length = (STRLEN) list[0];
7243 const UV version_id = list[1];
7244 const bool offset = cBOOL(list[2]);
7245 #define HEADER_LENGTH 3
7246 /* If any of the above changes in any way, you must change HEADER_LENGTH
7247 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
7248 * perl -E 'say int(rand 2**31-1)'
7250 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
7251 data structure type, so that one being
7252 passed in can be validated to be an
7253 inversion list of the correct vintage.
7256 SV* invlist = newSV_type(SVt_INVLIST);
7258 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7260 if (version_id != INVLIST_VERSION_ID) {
7261 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7264 /* The generated array passed in includes header elements that aren't part
7265 * of the list proper, so start it just after them */
7266 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
7268 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7269 shouldn't touch it */
7271 *(get_invlist_offset_addr(invlist)) = offset;
7273 /* The 'length' passed to us is the physical number of elements in the
7274 * inversion list. But if there is an offset the logical number is one
7276 invlist_set_len(invlist, length - offset, offset);
7278 invlist_set_previous_index(invlist, 0);
7280 /* Initialize the iteration pointer. */
7281 invlist_iterfinish(invlist);
7287 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7289 /* Grow the maximum size of an inversion list */
7291 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7293 assert(SvTYPE(invlist) == SVt_INVLIST);
7295 /* Add one to account for the zero element at the beginning which may not
7296 * be counted by the calling parameters */
7297 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
7300 PERL_STATIC_INLINE void
7301 S_invlist_trim(pTHX_ SV* const invlist)
7303 PERL_ARGS_ASSERT_INVLIST_TRIM;
7305 assert(SvTYPE(invlist) == SVt_INVLIST);
7307 /* Change the length of the inversion list to how many entries it currently
7309 SvPV_shrink_to_cur((SV *) invlist);
7312 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7315 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7317 /* Subject to change or removal. Append the range from 'start' to 'end' at
7318 * the end of the inversion list. The range must be above any existing
7322 UV max = invlist_max(invlist);
7323 UV len = _invlist_len(invlist);
7326 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7328 if (len == 0) { /* Empty lists must be initialized */
7329 offset = start != 0;
7330 array = _invlist_array_init(invlist, ! offset);
7333 /* Here, the existing list is non-empty. The current max entry in the
7334 * list is generally the first value not in the set, except when the
7335 * set extends to the end of permissible values, in which case it is
7336 * the first entry in that final set, and so this call is an attempt to
7337 * append out-of-order */
7339 UV final_element = len - 1;
7340 array = invlist_array(invlist);
7341 if (array[final_element] > start
7342 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7344 Perl_croak(aTHX_ "panic: attempting to append to an inversion list, but wasn't at the end of the list, final=%"UVuf", start=%"UVuf", match=%c",
7345 array[final_element], start,
7346 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7349 /* Here, it is a legal append. If the new range begins with the first
7350 * value not in the set, it is extending the set, so the new first
7351 * value not in the set is one greater than the newly extended range.
7353 offset = *get_invlist_offset_addr(invlist);
7354 if (array[final_element] == start) {
7355 if (end != UV_MAX) {
7356 array[final_element] = end + 1;
7359 /* But if the end is the maximum representable on the machine,
7360 * just let the range that this would extend to have no end */
7361 invlist_set_len(invlist, len - 1, offset);
7367 /* Here the new range doesn't extend any existing set. Add it */
7369 len += 2; /* Includes an element each for the start and end of range */
7371 /* If wll overflow the existing space, extend, which may cause the array to
7374 invlist_extend(invlist, len);
7376 /* Have to set len here to avoid assert failure in invlist_array() */
7377 invlist_set_len(invlist, len, offset);
7379 array = invlist_array(invlist);
7382 invlist_set_len(invlist, len, offset);
7385 /* The next item on the list starts the range, the one after that is
7386 * one past the new range. */
7387 array[len - 2] = start;
7388 if (end != UV_MAX) {
7389 array[len - 1] = end + 1;
7392 /* But if the end is the maximum representable on the machine, just let
7393 * the range have no end */
7394 invlist_set_len(invlist, len - 1, offset);
7398 #ifndef PERL_IN_XSUB_RE
7401 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7403 /* Searches the inversion list for the entry that contains the input code
7404 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7405 * return value is the index into the list's array of the range that
7410 IV high = _invlist_len(invlist);
7411 const IV highest_element = high - 1;
7414 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7416 /* If list is empty, return failure. */
7421 /* (We can't get the array unless we know the list is non-empty) */
7422 array = invlist_array(invlist);
7424 mid = invlist_previous_index(invlist);
7425 assert(mid >=0 && mid <= highest_element);
7427 /* <mid> contains the cache of the result of the previous call to this
7428 * function (0 the first time). See if this call is for the same result,
7429 * or if it is for mid-1. This is under the theory that calls to this
7430 * function will often be for related code points that are near each other.
7431 * And benchmarks show that caching gives better results. We also test
7432 * here if the code point is within the bounds of the list. These tests
7433 * replace others that would have had to be made anyway to make sure that
7434 * the array bounds were not exceeded, and these give us extra information
7435 * at the same time */
7436 if (cp >= array[mid]) {
7437 if (cp >= array[highest_element]) {
7438 return highest_element;
7441 /* Here, array[mid] <= cp < array[highest_element]. This means that
7442 * the final element is not the answer, so can exclude it; it also
7443 * means that <mid> is not the final element, so can refer to 'mid + 1'
7445 if (cp < array[mid + 1]) {
7451 else { /* cp < aray[mid] */
7452 if (cp < array[0]) { /* Fail if outside the array */
7456 if (cp >= array[mid - 1]) {
7461 /* Binary search. What we are looking for is <i> such that
7462 * array[i] <= cp < array[i+1]
7463 * The loop below converges on the i+1. Note that there may not be an
7464 * (i+1)th element in the array, and things work nonetheless */
7465 while (low < high) {
7466 mid = (low + high) / 2;
7467 assert(mid <= highest_element);
7468 if (array[mid] <= cp) { /* cp >= array[mid] */
7471 /* We could do this extra test to exit the loop early.
7472 if (cp < array[low]) {
7477 else { /* cp < array[mid] */
7484 invlist_set_previous_index(invlist, high);
7489 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7491 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7492 * but is used when the swash has an inversion list. This makes this much
7493 * faster, as it uses a binary search instead of a linear one. This is
7494 * intimately tied to that function, and perhaps should be in utf8.c,
7495 * except it is intimately tied to inversion lists as well. It assumes
7496 * that <swatch> is all 0's on input */
7499 const IV len = _invlist_len(invlist);
7503 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7505 if (len == 0) { /* Empty inversion list */
7509 array = invlist_array(invlist);
7511 /* Find which element it is */
7512 i = _invlist_search(invlist, start);
7514 /* We populate from <start> to <end> */
7515 while (current < end) {
7518 /* The inversion list gives the results for every possible code point
7519 * after the first one in the list. Only those ranges whose index is
7520 * even are ones that the inversion list matches. For the odd ones,
7521 * and if the initial code point is not in the list, we have to skip
7522 * forward to the next element */
7523 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7525 if (i >= len) { /* Finished if beyond the end of the array */
7529 if (current >= end) { /* Finished if beyond the end of what we
7531 if (LIKELY(end < UV_MAX)) {
7535 /* We get here when the upper bound is the maximum
7536 * representable on the machine, and we are looking for just
7537 * that code point. Have to special case it */
7539 goto join_end_of_list;
7542 assert(current >= start);
7544 /* The current range ends one below the next one, except don't go past
7547 upper = (i < len && array[i] < end) ? array[i] : end;
7549 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7550 * for each code point in it */
7551 for (; current < upper; current++) {
7552 const STRLEN offset = (STRLEN)(current - start);
7553 swatch[offset >> 3] |= 1 << (offset & 7);
7558 /* Quit if at the end of the list */
7561 /* But first, have to deal with the highest possible code point on
7562 * the platform. The previous code assumes that <end> is one
7563 * beyond where we want to populate, but that is impossible at the
7564 * platform's infinity, so have to handle it specially */
7565 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7567 const STRLEN offset = (STRLEN)(end - start);
7568 swatch[offset >> 3] |= 1 << (offset & 7);
7573 /* Advance to the next range, which will be for code points not in the
7582 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, const bool complement_b, SV** output)
7584 /* Take the union of two inversion lists and point <output> to it. *output
7585 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7586 * the reference count to that list will be decremented. The first list,
7587 * <a>, may be NULL, in which case a copy of the second list is returned.
7588 * If <complement_b> is TRUE, the union is taken of the complement
7589 * (inversion) of <b> instead of b itself.
7591 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7592 * Richard Gillam, published by Addison-Wesley, and explained at some
7593 * length there. The preface says to incorporate its examples into your
7594 * code at your own risk.
7596 * The algorithm is like a merge sort.
7598 * XXX A potential performance improvement is to keep track as we go along
7599 * if only one of the inputs contributes to the result, meaning the other
7600 * is a subset of that one. In that case, we can skip the final copy and
7601 * return the larger of the input lists, but then outside code might need
7602 * to keep track of whether to free the input list or not */
7604 const UV* array_a; /* a's array */
7606 UV len_a; /* length of a's array */
7609 SV* u; /* the resulting union */
7613 UV i_a = 0; /* current index into a's array */
7617 /* running count, as explained in the algorithm source book; items are
7618 * stopped accumulating and are output when the count changes to/from 0.
7619 * The count is incremented when we start a range that's in the set, and
7620 * decremented when we start a range that's not in the set. So its range
7621 * is 0 to 2. Only when the count is zero is something not in the set.
7625 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7628 /* If either one is empty, the union is the other one */
7629 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7636 *output = invlist_clone(b);
7638 _invlist_invert(*output);
7640 } /* else *output already = b; */
7643 else if ((len_b = _invlist_len(b)) == 0) {
7648 /* The complement of an empty list is a list that has everything in it,
7649 * so the union with <a> includes everything too */
7654 *output = _new_invlist(1);
7655 _append_range_to_invlist(*output, 0, UV_MAX);
7657 else if (*output != a) {
7658 *output = invlist_clone(a);
7660 /* else *output already = a; */
7664 /* Here both lists exist and are non-empty */
7665 array_a = invlist_array(a);
7666 array_b = invlist_array(b);
7668 /* If are to take the union of 'a' with the complement of b, set it
7669 * up so are looking at b's complement. */
7672 /* To complement, we invert: if the first element is 0, remove it. To
7673 * do this, we just pretend the array starts one later */
7674 if (array_b[0] == 0) {
7680 /* But if the first element is not zero, we pretend the list starts
7681 * at the 0 that is always stored immediately before the array. */
7687 /* Size the union for the worst case: that the sets are completely
7689 u = _new_invlist(len_a + len_b);
7691 /* Will contain U+0000 if either component does */
7692 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7693 || (len_b > 0 && array_b[0] == 0));
7695 /* Go through each list item by item, stopping when exhausted one of
7697 while (i_a < len_a && i_b < len_b) {
7698 UV cp; /* The element to potentially add to the union's array */
7699 bool cp_in_set; /* is it in the the input list's set or not */
7701 /* We need to take one or the other of the two inputs for the union.
7702 * Since we are merging two sorted lists, we take the smaller of the
7703 * next items. In case of a tie, we take the one that is in its set
7704 * first. If we took one not in the set first, it would decrement the
7705 * count, possibly to 0 which would cause it to be output as ending the
7706 * range, and the next time through we would take the same number, and
7707 * output it again as beginning the next range. By doing it the
7708 * opposite way, there is no possibility that the count will be
7709 * momentarily decremented to 0, and thus the two adjoining ranges will
7710 * be seamlessly merged. (In a tie and both are in the set or both not
7711 * in the set, it doesn't matter which we take first.) */
7712 if (array_a[i_a] < array_b[i_b]
7713 || (array_a[i_a] == array_b[i_b]
7714 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7716 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7720 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7721 cp = array_b[i_b++];
7724 /* Here, have chosen which of the two inputs to look at. Only output
7725 * if the running count changes to/from 0, which marks the
7726 * beginning/end of a range in that's in the set */
7729 array_u[i_u++] = cp;
7736 array_u[i_u++] = cp;
7741 /* Here, we are finished going through at least one of the lists, which
7742 * means there is something remaining in at most one. We check if the list
7743 * that hasn't been exhausted is positioned such that we are in the middle
7744 * of a range in its set or not. (i_a and i_b point to the element beyond
7745 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7746 * is potentially more to output.
7747 * There are four cases:
7748 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7749 * in the union is entirely from the non-exhausted set.
7750 * 2) Both were in their sets, count is 2. Nothing further should
7751 * be output, as everything that remains will be in the exhausted
7752 * list's set, hence in the union; decrementing to 1 but not 0 insures
7754 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7755 * Nothing further should be output because the union includes
7756 * everything from the exhausted set. Not decrementing ensures that.
7757 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7758 * decrementing to 0 insures that we look at the remainder of the
7759 * non-exhausted set */
7760 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7761 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7766 /* The final length is what we've output so far, plus what else is about to
7767 * be output. (If 'count' is non-zero, then the input list we exhausted
7768 * has everything remaining up to the machine's limit in its set, and hence
7769 * in the union, so there will be no further output. */
7772 /* At most one of the subexpressions will be non-zero */
7773 len_u += (len_a - i_a) + (len_b - i_b);
7776 /* Set result to final length, which can change the pointer to array_u, so
7778 if (len_u != _invlist_len(u)) {
7779 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
7781 array_u = invlist_array(u);
7784 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7785 * the other) ended with everything above it not in its set. That means
7786 * that the remaining part of the union is precisely the same as the
7787 * non-exhausted list, so can just copy it unchanged. (If both list were
7788 * exhausted at the same time, then the operations below will be both 0.)
7791 IV copy_count; /* At most one will have a non-zero copy count */
7792 if ((copy_count = len_a - i_a) > 0) {
7793 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7795 else if ((copy_count = len_b - i_b) > 0) {
7796 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7800 /* We may be removing a reference to one of the inputs */
7801 if (a == *output || b == *output) {
7802 assert(! invlist_is_iterating(*output));
7803 SvREFCNT_dec_NN(*output);
7811 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, const bool complement_b, SV** i)
7813 /* Take the intersection of two inversion lists and point <i> to it. *i
7814 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7815 * the reference count to that list will be decremented.
7816 * If <complement_b> is TRUE, the result will be the intersection of <a>
7817 * and the complement (or inversion) of <b> instead of <b> directly.
7819 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7820 * Richard Gillam, published by Addison-Wesley, and explained at some
7821 * length there. The preface says to incorporate its examples into your
7822 * code at your own risk. In fact, it had bugs
7824 * The algorithm is like a merge sort, and is essentially the same as the
7828 const UV* array_a; /* a's array */
7830 UV len_a; /* length of a's array */
7833 SV* r; /* the resulting intersection */
7837 UV i_a = 0; /* current index into a's array */
7841 /* running count, as explained in the algorithm source book; items are
7842 * stopped accumulating and are output when the count changes to/from 2.
7843 * The count is incremented when we start a range that's in the set, and
7844 * decremented when we start a range that's not in the set. So its range
7845 * is 0 to 2. Only when the count is 2 is something in the intersection.
7849 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7852 /* Special case if either one is empty */
7853 len_a = (a == NULL) ? 0 : _invlist_len(a);
7854 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7856 if (len_a != 0 && complement_b) {
7858 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7859 * be empty. Here, also we are using 'b's complement, which hence
7860 * must be every possible code point. Thus the intersection is
7867 *i = invlist_clone(a);
7869 /* else *i is already 'a' */
7873 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7874 * intersection must be empty */
7881 *i = _new_invlist(0);
7885 /* Here both lists exist and are non-empty */
7886 array_a = invlist_array(a);
7887 array_b = invlist_array(b);
7889 /* If are to take the intersection of 'a' with the complement of b, set it
7890 * up so are looking at b's complement. */
7893 /* To complement, we invert: if the first element is 0, remove it. To
7894 * do this, we just pretend the array starts one later */
7895 if (array_b[0] == 0) {
7901 /* But if the first element is not zero, we pretend the list starts
7902 * at the 0 that is always stored immediately before the array. */
7908 /* Size the intersection for the worst case: that the intersection ends up
7909 * fragmenting everything to be completely disjoint */
7910 r= _new_invlist(len_a + len_b);
7912 /* Will contain U+0000 iff both components do */
7913 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7914 && len_b > 0 && array_b[0] == 0);
7916 /* Go through each list item by item, stopping when exhausted one of
7918 while (i_a < len_a && i_b < len_b) {
7919 UV cp; /* The element to potentially add to the intersection's
7921 bool cp_in_set; /* Is it in the input list's set or not */
7923 /* We need to take one or the other of the two inputs for the
7924 * intersection. Since we are merging two sorted lists, we take the
7925 * smaller of the next items. In case of a tie, we take the one that
7926 * is not in its set first (a difference from the union algorithm). If
7927 * we took one in the set first, it would increment the count, possibly
7928 * to 2 which would cause it to be output as starting a range in the
7929 * intersection, and the next time through we would take that same
7930 * number, and output it again as ending the set. By doing it the
7931 * opposite of this, there is no possibility that the count will be
7932 * momentarily incremented to 2. (In a tie and both are in the set or
7933 * both not in the set, it doesn't matter which we take first.) */
7934 if (array_a[i_a] < array_b[i_b]
7935 || (array_a[i_a] == array_b[i_b]
7936 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7938 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7942 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7946 /* Here, have chosen which of the two inputs to look at. Only output
7947 * if the running count changes to/from 2, which marks the
7948 * beginning/end of a range that's in the intersection */
7952 array_r[i_r++] = cp;
7957 array_r[i_r++] = cp;
7963 /* Here, we are finished going through at least one of the lists, which
7964 * means there is something remaining in at most one. We check if the list
7965 * that has been exhausted is positioned such that we are in the middle
7966 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7967 * the ones we care about.) There are four cases:
7968 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7969 * nothing left in the intersection.
7970 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7971 * above 2. What should be output is exactly that which is in the
7972 * non-exhausted set, as everything it has is also in the intersection
7973 * set, and everything it doesn't have can't be in the intersection
7974 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7975 * gets incremented to 2. Like the previous case, the intersection is
7976 * everything that remains in the non-exhausted set.
7977 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7978 * remains 1. And the intersection has nothing more. */
7979 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7980 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7985 /* The final length is what we've output so far plus what else is in the
7986 * intersection. At most one of the subexpressions below will be non-zero */
7989 len_r += (len_a - i_a) + (len_b - i_b);
7992 /* Set result to final length, which can change the pointer to array_r, so
7994 if (len_r != _invlist_len(r)) {
7995 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
7997 array_r = invlist_array(r);
8000 /* Finish outputting any remaining */
8001 if (count >= 2) { /* At most one will have a non-zero copy count */
8003 if ((copy_count = len_a - i_a) > 0) {
8004 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
8006 else if ((copy_count = len_b - i_b) > 0) {
8007 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8011 /* We may be removing a reference to one of the inputs */
8012 if (a == *i || b == *i) {
8013 assert(! invlist_is_iterating(*i));
8014 SvREFCNT_dec_NN(*i);
8022 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8024 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8025 * set. A pointer to the inversion list is returned. This may actually be
8026 * a new list, in which case the passed in one has been destroyed. The
8027 * passed in inversion list can be NULL, in which case a new one is created
8028 * with just the one range in it */
8033 if (invlist == NULL) {
8034 invlist = _new_invlist(2);
8038 len = _invlist_len(invlist);
8041 /* If comes after the final entry actually in the list, can just append it
8044 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8045 && start >= invlist_array(invlist)[len - 1]))
8047 _append_range_to_invlist(invlist, start, end);
8051 /* Here, can't just append things, create and return a new inversion list
8052 * which is the union of this range and the existing inversion list */
8053 range_invlist = _new_invlist(2);
8054 _append_range_to_invlist(range_invlist, start, end);
8056 _invlist_union(invlist, range_invlist, &invlist);
8058 /* The temporary can be freed */
8059 SvREFCNT_dec_NN(range_invlist);
8066 PERL_STATIC_INLINE SV*
8067 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8068 return _add_range_to_invlist(invlist, cp, cp);
8071 #ifndef PERL_IN_XSUB_RE
8073 Perl__invlist_invert(pTHX_ SV* const invlist)
8075 /* Complement the input inversion list. This adds a 0 if the list didn't
8076 * have a zero; removes it otherwise. As described above, the data
8077 * structure is set up so that this is very efficient */
8079 PERL_ARGS_ASSERT__INVLIST_INVERT;
8081 assert(! invlist_is_iterating(invlist));
8083 /* The inverse of matching nothing is matching everything */
8084 if (_invlist_len(invlist) == 0) {
8085 _append_range_to_invlist(invlist, 0, UV_MAX);
8089 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
8093 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
8095 /* Complement the input inversion list (which must be a Unicode property,
8096 * all of which don't match above the Unicode maximum code point.) And
8097 * Perl has chosen to not have the inversion match above that either. This
8098 * adds a 0x110000 if the list didn't end with it, and removes it if it did
8104 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
8106 _invlist_invert(invlist);
8108 len = _invlist_len(invlist);
8110 if (len != 0) { /* If empty do nothing */
8111 array = invlist_array(invlist);
8112 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8113 /* Add 0x110000. First, grow if necessary */
8115 if (invlist_max(invlist) < len) {
8116 invlist_extend(invlist, len);
8117 array = invlist_array(invlist);
8119 invlist_set_len(invlist, len, *get_invlist_offset_addr(invlist));
8120 array[len - 1] = PERL_UNICODE_MAX + 1;
8122 else { /* Remove the 0x110000 */
8123 invlist_set_len(invlist, len - 1, *get_invlist_offset_addr(invlist));
8131 PERL_STATIC_INLINE SV*
8132 S_invlist_clone(pTHX_ SV* const invlist)
8135 /* Return a new inversion list that is a copy of the input one, which is
8138 /* Need to allocate extra space to accommodate Perl's addition of a
8139 * trailing NUL to SvPV's, since it thinks they are always strings */
8140 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8141 STRLEN physical_length = SvCUR(invlist);
8142 bool offset = *(get_invlist_offset_addr(invlist));
8144 PERL_ARGS_ASSERT_INVLIST_CLONE;
8146 *(get_invlist_offset_addr(new_invlist)) = offset;
8147 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
8148 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
8153 PERL_STATIC_INLINE STRLEN*
8154 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8156 /* Return the address of the UV that contains the current iteration
8159 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8161 assert(SvTYPE(invlist) == SVt_INVLIST);
8163 return &(((XINVLIST*) SvANY(invlist))->iterator);
8166 PERL_STATIC_INLINE void
8167 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8169 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8171 *get_invlist_iter_addr(invlist) = 0;
8174 PERL_STATIC_INLINE void
8175 S_invlist_iterfinish(pTHX_ SV* invlist)
8177 /* Terminate iterator for invlist. This is to catch development errors.
8178 * Any iteration that is interrupted before completed should call this
8179 * function. Functions that add code points anywhere else but to the end
8180 * of an inversion list assert that they are not in the middle of an
8181 * iteration. If they were, the addition would make the iteration
8182 * problematical: if the iteration hadn't reached the place where things
8183 * were being added, it would be ok */
8185 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8187 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
8191 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8193 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8194 * This call sets in <*start> and <*end>, the next range in <invlist>.
8195 * Returns <TRUE> if successful and the next call will return the next
8196 * range; <FALSE> if was already at the end of the list. If the latter,
8197 * <*start> and <*end> are unchanged, and the next call to this function
8198 * will start over at the beginning of the list */
8200 STRLEN* pos = get_invlist_iter_addr(invlist);
8201 UV len = _invlist_len(invlist);
8204 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8207 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
8211 array = invlist_array(invlist);
8213 *start = array[(*pos)++];
8219 *end = array[(*pos)++] - 1;
8225 PERL_STATIC_INLINE bool
8226 S_invlist_is_iterating(pTHX_ SV* const invlist)
8228 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8230 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8233 PERL_STATIC_INLINE UV
8234 S_invlist_highest(pTHX_ SV* const invlist)
8236 /* Returns the highest code point that matches an inversion list. This API
8237 * has an ambiguity, as it returns 0 under either the highest is actually
8238 * 0, or if the list is empty. If this distinction matters to you, check
8239 * for emptiness before calling this function */
8241 UV len = _invlist_len(invlist);
8244 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8250 array = invlist_array(invlist);
8252 /* The last element in the array in the inversion list always starts a
8253 * range that goes to infinity. That range may be for code points that are
8254 * matched in the inversion list, or it may be for ones that aren't
8255 * matched. In the latter case, the highest code point in the set is one
8256 * less than the beginning of this range; otherwise it is the final element
8257 * of this range: infinity */
8258 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8260 : array[len - 1] - 1;
8263 #ifndef PERL_IN_XSUB_RE
8265 Perl__invlist_contents(pTHX_ SV* const invlist)
8267 /* Get the contents of an inversion list into a string SV so that they can
8268 * be printed out. It uses the format traditionally done for debug tracing
8272 SV* output = newSVpvs("\n");
8274 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8276 assert(! invlist_is_iterating(invlist));
8278 invlist_iterinit(invlist);
8279 while (invlist_iternext(invlist, &start, &end)) {
8280 if (end == UV_MAX) {
8281 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8283 else if (end != start) {
8284 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8288 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8296 #ifndef PERL_IN_XSUB_RE
8298 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level, const char * const indent, SV* const invlist)
8300 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
8301 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
8302 * the string 'indent'. The output looks like this:
8303 [0] 0x000A .. 0x000D
8305 [4] 0x2028 .. 0x2029
8306 [6] 0x3104 .. INFINITY
8307 * This means that the first range of code points matched by the list are
8308 * 0xA through 0xD; the second range contains only the single code point
8309 * 0x85, etc. An inversion list is an array of UVs. Two array elements
8310 * are used to define each range (except if the final range extends to
8311 * infinity, only a single element is needed). The array index of the
8312 * first element for the corresponding range is given in brackets. */
8317 PERL_ARGS_ASSERT__INVLIST_DUMP;
8319 if (invlist_is_iterating(invlist)) {
8320 Perl_dump_indent(aTHX_ level, file,
8321 "%sCan't dump inversion list because is in middle of iterating\n",
8326 invlist_iterinit(invlist);
8327 while (invlist_iternext(invlist, &start, &end)) {
8328 if (end == UV_MAX) {
8329 Perl_dump_indent(aTHX_ level, file,
8330 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
8331 indent, (UV)count, start);
8333 else if (end != start) {
8334 Perl_dump_indent(aTHX_ level, file,
8335 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
8336 indent, (UV)count, start, end);
8339 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
8340 indent, (UV)count, start);
8347 #ifdef PERL_ARGS_ASSERT__INVLISTEQ
8349 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
8351 /* Return a boolean as to if the two passed in inversion lists are
8352 * identical. The final argument, if TRUE, says to take the complement of
8353 * the second inversion list before doing the comparison */
8355 const UV* array_a = invlist_array(a);
8356 const UV* array_b = invlist_array(b);
8357 UV len_a = _invlist_len(a);
8358 UV len_b = _invlist_len(b);
8360 UV i = 0; /* current index into the arrays */
8361 bool retval = TRUE; /* Assume are identical until proven otherwise */
8363 PERL_ARGS_ASSERT__INVLISTEQ;
8365 /* If are to compare 'a' with the complement of b, set it
8366 * up so are looking at b's complement. */
8369 /* The complement of nothing is everything, so <a> would have to have
8370 * just one element, starting at zero (ending at infinity) */
8372 return (len_a == 1 && array_a[0] == 0);
8374 else if (array_b[0] == 0) {
8376 /* Otherwise, to complement, we invert. Here, the first element is
8377 * 0, just remove it. To do this, we just pretend the array starts
8385 /* But if the first element is not zero, we pretend the list starts
8386 * at the 0 that is always stored immediately before the array. */
8392 /* Make sure that the lengths are the same, as well as the final element
8393 * before looping through the remainder. (Thus we test the length, final,
8394 * and first elements right off the bat) */
8395 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8398 else for (i = 0; i < len_a - 1; i++) {
8399 if (array_a[i] != array_b[i]) {
8409 #undef HEADER_LENGTH
8410 #undef TO_INTERNAL_SIZE
8411 #undef FROM_INTERNAL_SIZE
8412 #undef INVLIST_VERSION_ID
8414 /* End of inversion list object */
8417 S_parse_lparen_question_flags(pTHX_ struct RExC_state_t *pRExC_state)
8419 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
8420 * constructs, and updates RExC_flags with them. On input, RExC_parse
8421 * should point to the first flag; it is updated on output to point to the
8422 * final ')' or ':'. There needs to be at least one flag, or this will
8425 /* for (?g), (?gc), and (?o) warnings; warning
8426 about (?c) will warn about (?g) -- japhy */
8428 #define WASTED_O 0x01
8429 #define WASTED_G 0x02
8430 #define WASTED_C 0x04
8431 #define WASTED_GC (WASTED_G|WASTED_C)
8432 I32 wastedflags = 0x00;
8433 U32 posflags = 0, negflags = 0;
8434 U32 *flagsp = &posflags;
8435 char has_charset_modifier = '\0';
8437 bool has_use_defaults = FALSE;
8438 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
8440 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
8442 /* '^' as an initial flag sets certain defaults */
8443 if (UCHARAT(RExC_parse) == '^') {
8445 has_use_defaults = TRUE;
8446 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8447 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8448 ? REGEX_UNICODE_CHARSET
8449 : REGEX_DEPENDS_CHARSET);
8452 cs = get_regex_charset(RExC_flags);
8453 if (cs == REGEX_DEPENDS_CHARSET
8454 && (RExC_utf8 || RExC_uni_semantics))
8456 cs = REGEX_UNICODE_CHARSET;
8459 while (*RExC_parse) {
8460 /* && strchr("iogcmsx", *RExC_parse) */
8461 /* (?g), (?gc) and (?o) are useless here
8462 and must be globally applied -- japhy */
8463 switch (*RExC_parse) {
8465 /* Code for the imsx flags */
8466 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8468 case LOCALE_PAT_MOD:
8469 if (has_charset_modifier) {
8470 goto excess_modifier;
8472 else if (flagsp == &negflags) {
8475 cs = REGEX_LOCALE_CHARSET;
8476 has_charset_modifier = LOCALE_PAT_MOD;
8477 RExC_contains_locale = 1;
8479 case UNICODE_PAT_MOD:
8480 if (has_charset_modifier) {
8481 goto excess_modifier;
8483 else if (flagsp == &negflags) {
8486 cs = REGEX_UNICODE_CHARSET;
8487 has_charset_modifier = UNICODE_PAT_MOD;
8489 case ASCII_RESTRICT_PAT_MOD:
8490 if (flagsp == &negflags) {
8493 if (has_charset_modifier) {
8494 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8495 goto excess_modifier;
8497 /* Doubled modifier implies more restricted */
8498 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8501 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8503 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8505 case DEPENDS_PAT_MOD:
8506 if (has_use_defaults) {
8507 goto fail_modifiers;
8509 else if (flagsp == &negflags) {
8512 else if (has_charset_modifier) {
8513 goto excess_modifier;
8516 /* The dual charset means unicode semantics if the
8517 * pattern (or target, not known until runtime) are
8518 * utf8, or something in the pattern indicates unicode
8520 cs = (RExC_utf8 || RExC_uni_semantics)
8521 ? REGEX_UNICODE_CHARSET
8522 : REGEX_DEPENDS_CHARSET;
8523 has_charset_modifier = DEPENDS_PAT_MOD;
8527 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
8528 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
8530 else if (has_charset_modifier == *(RExC_parse - 1)) {
8531 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
8534 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
8539 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
8541 case ONCE_PAT_MOD: /* 'o' */
8542 case GLOBAL_PAT_MOD: /* 'g' */
8543 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8544 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
8545 if (! (wastedflags & wflagbit) ) {
8546 wastedflags |= wflagbit;
8547 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8550 "Useless (%s%c) - %suse /%c modifier",
8551 flagsp == &negflags ? "?-" : "?",
8553 flagsp == &negflags ? "don't " : "",
8560 case CONTINUE_PAT_MOD: /* 'c' */
8561 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8562 if (! (wastedflags & WASTED_C) ) {
8563 wastedflags |= WASTED_GC;
8564 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8567 "Useless (%sc) - %suse /gc modifier",
8568 flagsp == &negflags ? "?-" : "?",
8569 flagsp == &negflags ? "don't " : ""
8574 case KEEPCOPY_PAT_MOD: /* 'p' */
8575 if (flagsp == &negflags) {
8577 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
8579 *flagsp |= RXf_PMf_KEEPCOPY;
8583 /* A flag is a default iff it is following a minus, so
8584 * if there is a minus, it means will be trying to
8585 * re-specify a default which is an error */
8586 if (has_use_defaults || flagsp == &negflags) {
8587 goto fail_modifiers;
8590 wastedflags = 0; /* reset so (?g-c) warns twice */
8594 RExC_flags |= posflags;
8595 RExC_flags &= ~negflags;
8596 set_regex_charset(&RExC_flags, cs);
8602 vFAIL3("Sequence (%.*s...) not recognized",
8603 RExC_parse-seqstart, seqstart);
8612 - reg - regular expression, i.e. main body or parenthesized thing
8614 * Caller must absorb opening parenthesis.
8616 * Combining parenthesis handling with the base level of regular expression
8617 * is a trifle forced, but the need to tie the tails of the branches to what
8618 * follows makes it hard to avoid.
8620 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8622 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8624 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8627 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
8628 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
8629 needs to be restarted.
8630 Otherwise would only return NULL if regbranch() returns NULL, which
8633 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8634 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
8635 * 2 is like 1, but indicates that nextchar() has been called to advance
8636 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
8637 * this flag alerts us to the need to check for that */
8640 regnode *ret; /* Will be the head of the group. */
8643 regnode *ender = NULL;
8646 U32 oregflags = RExC_flags;
8647 bool have_branch = 0;
8649 I32 freeze_paren = 0;
8650 I32 after_freeze = 0;
8652 char * parse_start = RExC_parse; /* MJD */
8653 char * const oregcomp_parse = RExC_parse;
8655 GET_RE_DEBUG_FLAGS_DECL;
8657 PERL_ARGS_ASSERT_REG;
8658 DEBUG_PARSE("reg ");
8660 *flagp = 0; /* Tentatively. */
8663 /* Make an OPEN node, if parenthesized. */
8666 /* Under /x, space and comments can be gobbled up between the '(' and
8667 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
8668 * intervening space, as the sequence is a token, and a token should be
8670 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
8672 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8673 char *start_verb = RExC_parse;
8674 STRLEN verb_len = 0;
8675 char *start_arg = NULL;
8676 unsigned char op = 0;
8678 int internal_argval = 0; /* internal_argval is only useful if !argok */
8680 if (has_intervening_patws && SIZE_ONLY) {
8681 ckWARNregdep(RExC_parse + 1, "In '(*VERB...)', splitting the initial '(*' is deprecated");
8683 while ( *RExC_parse && *RExC_parse != ')' ) {
8684 if ( *RExC_parse == ':' ) {
8685 start_arg = RExC_parse + 1;
8691 verb_len = RExC_parse - start_verb;
8694 while ( *RExC_parse && *RExC_parse != ')' )
8696 if ( *RExC_parse != ')' )
8697 vFAIL("Unterminated verb pattern argument");
8698 if ( RExC_parse == start_arg )
8701 if ( *RExC_parse != ')' )
8702 vFAIL("Unterminated verb pattern");
8705 switch ( *start_verb ) {
8706 case 'A': /* (*ACCEPT) */
8707 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8709 internal_argval = RExC_nestroot;
8712 case 'C': /* (*COMMIT) */
8713 if ( memEQs(start_verb,verb_len,"COMMIT") )
8716 case 'F': /* (*FAIL) */
8717 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8722 case ':': /* (*:NAME) */
8723 case 'M': /* (*MARK:NAME) */
8724 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8729 case 'P': /* (*PRUNE) */
8730 if ( memEQs(start_verb,verb_len,"PRUNE") )
8733 case 'S': /* (*SKIP) */
8734 if ( memEQs(start_verb,verb_len,"SKIP") )
8737 case 'T': /* (*THEN) */
8738 /* [19:06] <TimToady> :: is then */
8739 if ( memEQs(start_verb,verb_len,"THEN") ) {
8741 RExC_seen |= REG_SEEN_CUTGROUP;
8747 vFAIL3("Unknown verb pattern '%.*s'",
8748 verb_len, start_verb);
8751 if ( start_arg && internal_argval ) {
8752 vFAIL3("Verb pattern '%.*s' may not have an argument",
8753 verb_len, start_verb);
8754 } else if ( argok < 0 && !start_arg ) {
8755 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8756 verb_len, start_verb);
8758 ret = reganode(pRExC_state, op, internal_argval);
8759 if ( ! internal_argval && ! SIZE_ONLY ) {
8761 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8762 ARG(ret) = add_data( pRExC_state, 1, "S" );
8763 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8770 if (!internal_argval)
8771 RExC_seen |= REG_SEEN_VERBARG;
8772 } else if ( start_arg ) {
8773 vFAIL3("Verb pattern '%.*s' may not have an argument",
8774 verb_len, start_verb);
8776 ret = reg_node(pRExC_state, op);
8778 nextchar(pRExC_state);
8781 else if (*RExC_parse == '?') { /* (?...) */
8782 bool is_logical = 0;
8783 const char * const seqstart = RExC_parse;
8784 if (has_intervening_patws && SIZE_ONLY) {
8785 ckWARNregdep(RExC_parse + 1, "In '(?...)', splitting the initial '(?' is deprecated");
8789 paren = *RExC_parse++;
8790 ret = NULL; /* For look-ahead/behind. */
8793 case 'P': /* (?P...) variants for those used to PCRE/Python */
8794 paren = *RExC_parse++;
8795 if ( paren == '<') /* (?P<...>) named capture */
8797 else if (paren == '>') { /* (?P>name) named recursion */
8798 goto named_recursion;
8800 else if (paren == '=') { /* (?P=...) named backref */
8801 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8802 you change this make sure you change that */
8803 char* name_start = RExC_parse;
8805 SV *sv_dat = reg_scan_name(pRExC_state,
8806 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8807 if (RExC_parse == name_start || *RExC_parse != ')')
8808 vFAIL2("Sequence %.3s... not terminated",parse_start);
8811 num = add_data( pRExC_state, 1, "S" );
8812 RExC_rxi->data->data[num]=(void*)sv_dat;
8813 SvREFCNT_inc_simple_void(sv_dat);
8816 ret = reganode(pRExC_state,
8819 : (ASCII_FOLD_RESTRICTED)
8821 : (AT_LEAST_UNI_SEMANTICS)
8829 Set_Node_Offset(ret, parse_start+1);
8830 Set_Node_Cur_Length(ret, parse_start);
8832 nextchar(pRExC_state);
8836 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8838 case '<': /* (?<...) */
8839 if (*RExC_parse == '!')
8841 else if (*RExC_parse != '=')
8847 case '\'': /* (?'...') */
8848 name_start= RExC_parse;
8849 svname = reg_scan_name(pRExC_state,
8850 SIZE_ONLY ? /* reverse test from the others */
8851 REG_RSN_RETURN_NAME :
8852 REG_RSN_RETURN_NULL);
8853 if (RExC_parse == name_start) {
8855 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8858 if (*RExC_parse != paren)
8859 vFAIL2("Sequence (?%c... not terminated",
8860 paren=='>' ? '<' : paren);
8864 if (!svname) /* shouldn't happen */
8866 "panic: reg_scan_name returned NULL");
8867 if (!RExC_paren_names) {
8868 RExC_paren_names= newHV();
8869 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8871 RExC_paren_name_list= newAV();
8872 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8875 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8877 sv_dat = HeVAL(he_str);
8879 /* croak baby croak */
8881 "panic: paren_name hash element allocation failed");
8882 } else if ( SvPOK(sv_dat) ) {
8883 /* (?|...) can mean we have dupes so scan to check
8884 its already been stored. Maybe a flag indicating
8885 we are inside such a construct would be useful,
8886 but the arrays are likely to be quite small, so
8887 for now we punt -- dmq */
8888 IV count = SvIV(sv_dat);
8889 I32 *pv = (I32*)SvPVX(sv_dat);
8891 for ( i = 0 ; i < count ; i++ ) {
8892 if ( pv[i] == RExC_npar ) {
8898 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8899 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8900 pv[count] = RExC_npar;
8901 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8904 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8905 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8907 SvIV_set(sv_dat, 1);
8910 /* Yes this does cause a memory leak in debugging Perls */
8911 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8912 SvREFCNT_dec_NN(svname);
8915 /*sv_dump(sv_dat);*/
8917 nextchar(pRExC_state);
8919 goto capturing_parens;
8921 RExC_seen |= REG_SEEN_LOOKBEHIND;
8922 RExC_in_lookbehind++;
8924 case '=': /* (?=...) */
8925 RExC_seen_zerolen++;
8927 case '!': /* (?!...) */
8928 RExC_seen_zerolen++;
8929 if (*RExC_parse == ')') {
8930 ret=reg_node(pRExC_state, OPFAIL);
8931 nextchar(pRExC_state);
8935 case '|': /* (?|...) */
8936 /* branch reset, behave like a (?:...) except that
8937 buffers in alternations share the same numbers */
8939 after_freeze = freeze_paren = RExC_npar;
8941 case ':': /* (?:...) */
8942 case '>': /* (?>...) */
8944 case '$': /* (?$...) */
8945 case '@': /* (?@...) */
8946 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8948 case '#': /* (?#...) */
8949 /* XXX As soon as we disallow separating the '?' and '*' (by
8950 * spaces or (?#...) comment), it is believed that this case
8951 * will be unreachable and can be removed. See
8953 while (*RExC_parse && *RExC_parse != ')')
8955 if (*RExC_parse != ')')
8956 FAIL("Sequence (?#... not terminated");
8957 nextchar(pRExC_state);
8960 case '0' : /* (?0) */
8961 case 'R' : /* (?R) */
8962 if (*RExC_parse != ')')
8963 FAIL("Sequence (?R) not terminated");
8964 ret = reg_node(pRExC_state, GOSTART);
8965 *flagp |= POSTPONED;
8966 nextchar(pRExC_state);
8969 { /* named and numeric backreferences */
8971 case '&': /* (?&NAME) */
8972 parse_start = RExC_parse - 1;
8975 SV *sv_dat = reg_scan_name(pRExC_state,
8976 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8977 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8979 goto gen_recurse_regop;
8980 assert(0); /* NOT REACHED */
8982 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8984 vFAIL("Illegal pattern");
8986 goto parse_recursion;
8988 case '-': /* (?-1) */
8989 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8990 RExC_parse--; /* rewind to let it be handled later */
8994 case '1': case '2': case '3': case '4': /* (?1) */
8995 case '5': case '6': case '7': case '8': case '9':
8998 num = atoi(RExC_parse);
8999 parse_start = RExC_parse - 1; /* MJD */
9000 if (*RExC_parse == '-')
9002 while (isDIGIT(*RExC_parse))
9004 if (*RExC_parse!=')')
9005 vFAIL("Expecting close bracket");
9008 if ( paren == '-' ) {
9010 Diagram of capture buffer numbering.
9011 Top line is the normal capture buffer numbers
9012 Bottom line is the negative indexing as from
9016 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
9020 num = RExC_npar + num;
9023 vFAIL("Reference to nonexistent group");
9025 } else if ( paren == '+' ) {
9026 num = RExC_npar + num - 1;
9029 ret = reganode(pRExC_state, GOSUB, num);
9031 if (num > (I32)RExC_rx->nparens) {
9033 vFAIL("Reference to nonexistent group");
9035 ARG2L_SET( ret, RExC_recurse_count++);
9037 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9038 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
9042 RExC_seen |= REG_SEEN_RECURSE;
9043 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9044 Set_Node_Offset(ret, parse_start); /* MJD */
9046 *flagp |= POSTPONED;
9047 nextchar(pRExC_state);
9049 } /* named and numeric backreferences */
9050 assert(0); /* NOT REACHED */
9052 case '?': /* (??...) */
9054 if (*RExC_parse != '{') {
9056 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
9059 *flagp |= POSTPONED;
9060 paren = *RExC_parse++;
9062 case '{': /* (?{...}) */
9065 struct reg_code_block *cb;
9067 RExC_seen_zerolen++;
9069 if ( !pRExC_state->num_code_blocks
9070 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9071 || pRExC_state->code_blocks[pRExC_state->code_index].start
9072 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9075 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9076 FAIL("panic: Sequence (?{...}): no code block found\n");
9077 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9079 /* this is a pre-compiled code block (?{...}) */
9080 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9081 RExC_parse = RExC_start + cb->end;
9084 if (cb->src_regex) {
9085 n = add_data(pRExC_state, 2, "rl");
9086 RExC_rxi->data->data[n] =
9087 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9088 RExC_rxi->data->data[n+1] = (void*)o;
9091 n = add_data(pRExC_state, 1,
9092 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
9093 RExC_rxi->data->data[n] = (void*)o;
9096 pRExC_state->code_index++;
9097 nextchar(pRExC_state);
9101 ret = reg_node(pRExC_state, LOGICAL);
9102 eval = reganode(pRExC_state, EVAL, n);
9105 /* for later propagation into (??{}) return value */
9106 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9108 REGTAIL(pRExC_state, ret, eval);
9109 /* deal with the length of this later - MJD */
9112 ret = reganode(pRExC_state, EVAL, n);
9113 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9114 Set_Node_Offset(ret, parse_start);
9117 case '(': /* (?(?{...})...) and (?(?=...)...) */
9120 if (RExC_parse[0] == '?') { /* (?(?...)) */
9121 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9122 || RExC_parse[1] == '<'
9123 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9127 ret = reg_node(pRExC_state, LOGICAL);
9131 tail = reg(pRExC_state, 1, &flag, depth+1);
9132 if (flag & RESTART_UTF8) {
9133 *flagp = RESTART_UTF8;
9136 REGTAIL(pRExC_state, ret, tail);
9140 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9141 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9143 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9144 char *name_start= RExC_parse++;
9146 SV *sv_dat=reg_scan_name(pRExC_state,
9147 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9148 if (RExC_parse == name_start || *RExC_parse != ch)
9149 vFAIL2("Sequence (?(%c... not terminated",
9150 (ch == '>' ? '<' : ch));
9153 num = add_data( pRExC_state, 1, "S" );
9154 RExC_rxi->data->data[num]=(void*)sv_dat;
9155 SvREFCNT_inc_simple_void(sv_dat);
9157 ret = reganode(pRExC_state,NGROUPP,num);
9158 goto insert_if_check_paren;
9160 else if (RExC_parse[0] == 'D' &&
9161 RExC_parse[1] == 'E' &&
9162 RExC_parse[2] == 'F' &&
9163 RExC_parse[3] == 'I' &&
9164 RExC_parse[4] == 'N' &&
9165 RExC_parse[5] == 'E')
9167 ret = reganode(pRExC_state,DEFINEP,0);
9170 goto insert_if_check_paren;
9172 else if (RExC_parse[0] == 'R') {
9175 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9176 parno = atoi(RExC_parse++);
9177 while (isDIGIT(*RExC_parse))
9179 } else if (RExC_parse[0] == '&') {
9182 sv_dat = reg_scan_name(pRExC_state,
9183 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9184 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9186 ret = reganode(pRExC_state,INSUBP,parno);
9187 goto insert_if_check_paren;
9189 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9192 parno = atoi(RExC_parse++);
9194 while (isDIGIT(*RExC_parse))
9196 ret = reganode(pRExC_state, GROUPP, parno);
9198 insert_if_check_paren:
9199 if ((c = *nextchar(pRExC_state)) != ')')
9200 vFAIL("Switch condition not recognized");
9202 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9203 br = regbranch(pRExC_state, &flags, 1,depth+1);
9205 if (flags & RESTART_UTF8) {
9206 *flagp = RESTART_UTF8;
9209 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9212 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
9213 c = *nextchar(pRExC_state);
9218 vFAIL("(?(DEFINE)....) does not allow branches");
9219 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
9220 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9221 if (flags & RESTART_UTF8) {
9222 *flagp = RESTART_UTF8;
9225 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9228 REGTAIL(pRExC_state, ret, lastbr);
9231 c = *nextchar(pRExC_state);
9236 vFAIL("Switch (?(condition)... contains too many branches");
9237 ender = reg_node(pRExC_state, TAIL);
9238 REGTAIL(pRExC_state, br, ender);
9240 REGTAIL(pRExC_state, lastbr, ender);
9241 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9244 REGTAIL(pRExC_state, ret, ender);
9245 RExC_size++; /* XXX WHY do we need this?!!
9246 For large programs it seems to be required
9247 but I can't figure out why. -- dmq*/
9251 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
9254 case '[': /* (?[ ... ]) */
9255 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9258 RExC_parse--; /* for vFAIL to print correctly */
9259 vFAIL("Sequence (? incomplete");
9261 default: /* e.g., (?i) */
9264 parse_lparen_question_flags(pRExC_state);
9265 if (UCHARAT(RExC_parse) != ':') {
9266 nextchar(pRExC_state);
9271 nextchar(pRExC_state);
9281 ret = reganode(pRExC_state, OPEN, parno);
9284 RExC_nestroot = parno;
9285 if (RExC_seen & REG_SEEN_RECURSE
9286 && !RExC_open_parens[parno-1])
9288 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9289 "Setting open paren #%"IVdf" to %d\n",
9290 (IV)parno, REG_NODE_NUM(ret)));
9291 RExC_open_parens[parno-1]= ret;
9294 Set_Node_Length(ret, 1); /* MJD */
9295 Set_Node_Offset(ret, RExC_parse); /* MJD */
9303 /* Pick up the branches, linking them together. */
9304 parse_start = RExC_parse; /* MJD */
9305 br = regbranch(pRExC_state, &flags, 1,depth+1);
9307 /* branch_len = (paren != 0); */
9310 if (flags & RESTART_UTF8) {
9311 *flagp = RESTART_UTF8;
9314 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9316 if (*RExC_parse == '|') {
9317 if (!SIZE_ONLY && RExC_extralen) {
9318 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9321 reginsert(pRExC_state, BRANCH, br, depth+1);
9322 Set_Node_Length(br, paren != 0);
9323 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9327 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9329 else if (paren == ':') {
9330 *flagp |= flags&SIMPLE;
9332 if (is_open) { /* Starts with OPEN. */
9333 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9335 else if (paren != '?') /* Not Conditional */
9337 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9339 while (*RExC_parse == '|') {
9340 if (!SIZE_ONLY && RExC_extralen) {
9341 ender = reganode(pRExC_state, LONGJMP,0);
9342 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9345 RExC_extralen += 2; /* Account for LONGJMP. */
9346 nextchar(pRExC_state);
9348 if (RExC_npar > after_freeze)
9349 after_freeze = RExC_npar;
9350 RExC_npar = freeze_paren;
9352 br = regbranch(pRExC_state, &flags, 0, depth+1);
9355 if (flags & RESTART_UTF8) {
9356 *flagp = RESTART_UTF8;
9359 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9361 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9363 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9366 if (have_branch || paren != ':') {
9367 /* Make a closing node, and hook it on the end. */
9370 ender = reg_node(pRExC_state, TAIL);
9373 ender = reganode(pRExC_state, CLOSE, parno);
9374 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9375 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9376 "Setting close paren #%"IVdf" to %d\n",
9377 (IV)parno, REG_NODE_NUM(ender)));
9378 RExC_close_parens[parno-1]= ender;
9379 if (RExC_nestroot == parno)
9382 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9383 Set_Node_Length(ender,1); /* MJD */
9389 *flagp &= ~HASWIDTH;
9392 ender = reg_node(pRExC_state, SUCCEED);
9395 ender = reg_node(pRExC_state, END);
9397 assert(!RExC_opend); /* there can only be one! */
9402 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9403 SV * const mysv_val1=sv_newmortal();
9404 SV * const mysv_val2=sv_newmortal();
9405 DEBUG_PARSE_MSG("lsbr");
9406 regprop(RExC_rx, mysv_val1, lastbr);
9407 regprop(RExC_rx, mysv_val2, ender);
9408 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9409 SvPV_nolen_const(mysv_val1),
9410 (IV)REG_NODE_NUM(lastbr),
9411 SvPV_nolen_const(mysv_val2),
9412 (IV)REG_NODE_NUM(ender),
9413 (IV)(ender - lastbr)
9416 REGTAIL(pRExC_state, lastbr, ender);
9418 if (have_branch && !SIZE_ONLY) {
9421 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9423 /* Hook the tails of the branches to the closing node. */
9424 for (br = ret; br; br = regnext(br)) {
9425 const U8 op = PL_regkind[OP(br)];
9427 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9428 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9431 else if (op == BRANCHJ) {
9432 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9433 /* for now we always disable this optimisation * /
9434 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9440 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9441 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9442 SV * const mysv_val1=sv_newmortal();
9443 SV * const mysv_val2=sv_newmortal();
9444 DEBUG_PARSE_MSG("NADA");
9445 regprop(RExC_rx, mysv_val1, ret);
9446 regprop(RExC_rx, mysv_val2, ender);
9447 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9448 SvPV_nolen_const(mysv_val1),
9449 (IV)REG_NODE_NUM(ret),
9450 SvPV_nolen_const(mysv_val2),
9451 (IV)REG_NODE_NUM(ender),
9456 if (OP(ender) == TAIL) {
9461 for ( opt= br + 1; opt < ender ; opt++ )
9463 NEXT_OFF(br)= ender - br;
9471 static const char parens[] = "=!<,>";
9473 if (paren && (p = strchr(parens, paren))) {
9474 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9475 int flag = (p - parens) > 1;
9478 node = SUSPEND, flag = 0;
9479 reginsert(pRExC_state, node,ret, depth+1);
9480 Set_Node_Cur_Length(ret, parse_start);
9481 Set_Node_Offset(ret, parse_start + 1);
9483 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9487 /* Check for proper termination. */
9489 /* restore original flags, but keep (?p) */
9490 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
9491 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9492 RExC_parse = oregcomp_parse;
9493 vFAIL("Unmatched (");
9496 else if (!paren && RExC_parse < RExC_end) {
9497 if (*RExC_parse == ')') {
9499 vFAIL("Unmatched )");
9502 FAIL("Junk on end of regexp"); /* "Can't happen". */
9503 assert(0); /* NOTREACHED */
9506 if (RExC_in_lookbehind) {
9507 RExC_in_lookbehind--;
9509 if (after_freeze > RExC_npar)
9510 RExC_npar = after_freeze;
9515 - regbranch - one alternative of an | operator
9517 * Implements the concatenation operator.
9519 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9523 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9527 regnode *chain = NULL;
9529 I32 flags = 0, c = 0;
9530 GET_RE_DEBUG_FLAGS_DECL;
9532 PERL_ARGS_ASSERT_REGBRANCH;
9534 DEBUG_PARSE("brnc");
9539 if (!SIZE_ONLY && RExC_extralen)
9540 ret = reganode(pRExC_state, BRANCHJ,0);
9542 ret = reg_node(pRExC_state, BRANCH);
9543 Set_Node_Length(ret, 1);
9547 if (!first && SIZE_ONLY)
9548 RExC_extralen += 1; /* BRANCHJ */
9550 *flagp = WORST; /* Tentatively. */
9553 nextchar(pRExC_state);
9554 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9556 latest = regpiece(pRExC_state, &flags,depth+1);
9557 if (latest == NULL) {
9558 if (flags & TRYAGAIN)
9560 if (flags & RESTART_UTF8) {
9561 *flagp = RESTART_UTF8;
9564 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
9566 else if (ret == NULL)
9568 *flagp |= flags&(HASWIDTH|POSTPONED);
9569 if (chain == NULL) /* First piece. */
9570 *flagp |= flags&SPSTART;
9573 REGTAIL(pRExC_state, chain, latest);
9578 if (chain == NULL) { /* Loop ran zero times. */
9579 chain = reg_node(pRExC_state, NOTHING);
9584 *flagp |= flags&SIMPLE;
9591 - regpiece - something followed by possible [*+?]
9593 * Note that the branching code sequences used for ? and the general cases
9594 * of * and + are somewhat optimized: they use the same NOTHING node as
9595 * both the endmarker for their branch list and the body of the last branch.
9596 * It might seem that this node could be dispensed with entirely, but the
9597 * endmarker role is not redundant.
9599 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
9601 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9605 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9612 const char * const origparse = RExC_parse;
9614 I32 max = REG_INFTY;
9615 #ifdef RE_TRACK_PATTERN_OFFSETS
9618 const char *maxpos = NULL;
9620 /* Save the original in case we change the emitted regop to a FAIL. */
9621 regnode * const orig_emit = RExC_emit;
9623 GET_RE_DEBUG_FLAGS_DECL;
9625 PERL_ARGS_ASSERT_REGPIECE;
9627 DEBUG_PARSE("piec");
9629 ret = regatom(pRExC_state, &flags,depth+1);
9631 if (flags & (TRYAGAIN|RESTART_UTF8))
9632 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
9634 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
9640 if (op == '{' && regcurly(RExC_parse, FALSE)) {
9642 #ifdef RE_TRACK_PATTERN_OFFSETS
9643 parse_start = RExC_parse; /* MJD */
9645 next = RExC_parse + 1;
9646 while (isDIGIT(*next) || *next == ',') {
9655 if (*next == '}') { /* got one */
9659 min = atoi(RExC_parse);
9663 maxpos = RExC_parse;
9665 if (!max && *maxpos != '0')
9666 max = REG_INFTY; /* meaning "infinity" */
9667 else if (max >= REG_INFTY)
9668 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9670 nextchar(pRExC_state);
9671 if (max < min) { /* If can't match, warn and optimize to fail
9674 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9676 /* We can't back off the size because we have to reserve
9677 * enough space for all the things we are about to throw
9678 * away, but we can shrink it by the ammount we are about
9680 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9683 RExC_emit = orig_emit;
9685 ret = reg_node(pRExC_state, OPFAIL);
9690 if ((flags&SIMPLE)) {
9691 RExC_naughty += 2 + RExC_naughty / 2;
9692 reginsert(pRExC_state, CURLY, ret, depth+1);
9693 Set_Node_Offset(ret, parse_start+1); /* MJD */
9694 Set_Node_Cur_Length(ret, parse_start);
9697 regnode * const w = reg_node(pRExC_state, WHILEM);
9700 REGTAIL(pRExC_state, ret, w);
9701 if (!SIZE_ONLY && RExC_extralen) {
9702 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9703 reginsert(pRExC_state, NOTHING,ret, depth+1);
9704 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9706 reginsert(pRExC_state, CURLYX,ret, depth+1);
9708 Set_Node_Offset(ret, parse_start+1);
9709 Set_Node_Length(ret,
9710 op == '{' ? (RExC_parse - parse_start) : 1);
9712 if (!SIZE_ONLY && RExC_extralen)
9713 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9714 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9716 RExC_whilem_seen++, RExC_extralen += 3;
9717 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9726 ARG1_SET(ret, (U16)min);
9727 ARG2_SET(ret, (U16)max);
9739 #if 0 /* Now runtime fix should be reliable. */
9741 /* if this is reinstated, don't forget to put this back into perldiag:
9743 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9745 (F) The part of the regexp subject to either the * or + quantifier
9746 could match an empty string. The {#} shows in the regular
9747 expression about where the problem was discovered.
9751 if (!(flags&HASWIDTH) && op != '?')
9752 vFAIL("Regexp *+ operand could be empty");
9755 #ifdef RE_TRACK_PATTERN_OFFSETS
9756 parse_start = RExC_parse;
9758 nextchar(pRExC_state);
9760 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9762 if (op == '*' && (flags&SIMPLE)) {
9763 reginsert(pRExC_state, STAR, ret, depth+1);
9767 else if (op == '*') {
9771 else if (op == '+' && (flags&SIMPLE)) {
9772 reginsert(pRExC_state, PLUS, ret, depth+1);
9776 else if (op == '+') {
9780 else if (op == '?') {
9785 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9786 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
9787 ckWARN3reg(RExC_parse,
9788 "%.*s matches null string many times",
9789 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9791 (void)ReREFCNT_inc(RExC_rx_sv);
9794 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9795 nextchar(pRExC_state);
9796 reginsert(pRExC_state, MINMOD, ret, depth+1);
9797 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9800 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9802 nextchar(pRExC_state);
9803 ender = reg_node(pRExC_state, SUCCEED);
9804 REGTAIL(pRExC_state, ret, ender);
9805 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9807 ender = reg_node(pRExC_state, TAIL);
9808 REGTAIL(pRExC_state, ret, ender);
9811 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9813 vFAIL("Nested quantifiers");
9820 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
9821 const bool strict /* Apply stricter parsing rules? */
9825 /* This is expected to be called by a parser routine that has recognized '\N'
9826 and needs to handle the rest. RExC_parse is expected to point at the first
9827 char following the N at the time of the call. On successful return,
9828 RExC_parse has been updated to point to just after the sequence identified
9829 by this routine, and <*flagp> has been updated.
9831 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9834 \N may begin either a named sequence, or if outside a character class, mean
9835 to match a non-newline. For non single-quoted regexes, the tokenizer has
9836 attempted to decide which, and in the case of a named sequence, converted it
9837 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9838 where c1... are the characters in the sequence. For single-quoted regexes,
9839 the tokenizer passes the \N sequence through unchanged; this code will not
9840 attempt to determine this nor expand those, instead raising a syntax error.
9841 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9842 or there is no '}', it signals that this \N occurrence means to match a
9845 Only the \N{U+...} form should occur in a character class, for the same
9846 reason that '.' inside a character class means to just match a period: it
9847 just doesn't make sense.
9849 The function raises an error (via vFAIL), and doesn't return for various
9850 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9851 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
9852 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
9853 only possible if node_p is non-NULL.
9856 If <valuep> is non-null, it means the caller can accept an input sequence
9857 consisting of a just a single code point; <*valuep> is set to that value
9858 if the input is such.
9860 If <node_p> is non-null it signifies that the caller can accept any other
9861 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9863 1) \N means not-a-NL: points to a newly created REG_ANY node;
9864 2) \N{}: points to a new NOTHING node;
9865 3) otherwise: points to a new EXACT node containing the resolved
9867 Note that FALSE is returned for single code point sequences if <valuep> is
9871 char * endbrace; /* '}' following the name */
9873 char *endchar; /* Points to '.' or '}' ending cur char in the input
9875 bool has_multiple_chars; /* true if the input stream contains a sequence of
9876 more than one character */
9878 GET_RE_DEBUG_FLAGS_DECL;
9880 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9884 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9886 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9887 * modifier. The other meaning does not */
9888 p = (RExC_flags & RXf_PMf_EXTENDED)
9889 ? regwhite( pRExC_state, RExC_parse )
9892 /* Disambiguate between \N meaning a named character versus \N meaning
9893 * [^\n]. The former is assumed when it can't be the latter. */
9894 if (*p != '{' || regcurly(p, FALSE)) {
9897 /* no bare \N in a charclass */
9898 if (in_char_class) {
9899 vFAIL("\\N in a character class must be a named character: \\N{...}");
9903 nextchar(pRExC_state);
9904 *node_p = reg_node(pRExC_state, REG_ANY);
9905 *flagp |= HASWIDTH|SIMPLE;
9908 Set_Node_Length(*node_p, 1); /* MJD */
9912 /* Here, we have decided it should be a named character or sequence */
9914 /* The test above made sure that the next real character is a '{', but
9915 * under the /x modifier, it could be separated by space (or a comment and
9916 * \n) and this is not allowed (for consistency with \x{...} and the
9917 * tokenizer handling of \N{NAME}). */
9918 if (*RExC_parse != '{') {
9919 vFAIL("Missing braces on \\N{}");
9922 RExC_parse++; /* Skip past the '{' */
9924 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9925 || ! (endbrace == RExC_parse /* nothing between the {} */
9926 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9927 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9929 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9930 vFAIL("\\N{NAME} must be resolved by the lexer");
9933 if (endbrace == RExC_parse) { /* empty: \N{} */
9936 *node_p = reg_node(pRExC_state,NOTHING);
9938 else if (in_char_class) {
9939 if (SIZE_ONLY && in_char_class) {
9941 RExC_parse++; /* Position after the "}" */
9942 vFAIL("Zero length \\N{}");
9945 ckWARNreg(RExC_parse,
9946 "Ignoring zero length \\N{} in character class");
9954 nextchar(pRExC_state);
9958 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9959 RExC_parse += 2; /* Skip past the 'U+' */
9961 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9963 /* Code points are separated by dots. If none, there is only one code
9964 * point, and is terminated by the brace */
9965 has_multiple_chars = (endchar < endbrace);
9967 if (valuep && (! has_multiple_chars || in_char_class)) {
9968 /* We only pay attention to the first char of
9969 multichar strings being returned in char classes. I kinda wonder
9970 if this makes sense as it does change the behaviour
9971 from earlier versions, OTOH that behaviour was broken
9972 as well. XXX Solution is to recharacterize as
9973 [rest-of-class]|multi1|multi2... */
9975 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9976 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9977 | PERL_SCAN_DISALLOW_PREFIX
9978 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9980 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9982 /* The tokenizer should have guaranteed validity, but it's possible to
9983 * bypass it by using single quoting, so check */
9984 if (length_of_hex == 0
9985 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9987 RExC_parse += length_of_hex; /* Includes all the valid */
9988 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
9989 ? UTF8SKIP(RExC_parse)
9991 /* Guard against malformed utf8 */
9992 if (RExC_parse >= endchar) {
9993 RExC_parse = endchar;
9995 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9998 if (in_char_class && has_multiple_chars) {
10000 RExC_parse = endbrace;
10001 vFAIL("\\N{} in character class restricted to one character");
10004 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
10008 RExC_parse = endbrace + 1;
10010 else if (! node_p || ! has_multiple_chars) {
10012 /* Here, the input is legal, but not according to the caller's
10013 * options. We fail without advancing the parse, so that the
10014 * caller can try again */
10020 /* What is done here is to convert this to a sub-pattern of the form
10021 * (?:\x{char1}\x{char2}...)
10022 * and then call reg recursively. That way, it retains its atomicness,
10023 * while not having to worry about special handling that some code
10024 * points may have. toke.c has converted the original Unicode values
10025 * to native, so that we can just pass on the hex values unchanged. We
10026 * do have to set a flag to keep recoding from happening in the
10029 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
10031 char *orig_end = RExC_end;
10034 while (RExC_parse < endbrace) {
10036 /* Convert to notation the rest of the code understands */
10037 sv_catpv(substitute_parse, "\\x{");
10038 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10039 sv_catpv(substitute_parse, "}");
10041 /* Point to the beginning of the next character in the sequence. */
10042 RExC_parse = endchar + 1;
10043 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10045 sv_catpv(substitute_parse, ")");
10047 RExC_parse = SvPV(substitute_parse, len);
10049 /* Don't allow empty number */
10051 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10053 RExC_end = RExC_parse + len;
10055 /* The values are Unicode, and therefore not subject to recoding */
10056 RExC_override_recoding = 1;
10058 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10059 if (flags & RESTART_UTF8) {
10060 *flagp = RESTART_UTF8;
10063 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
10066 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10068 RExC_parse = endbrace;
10069 RExC_end = orig_end;
10070 RExC_override_recoding = 0;
10072 nextchar(pRExC_state);
10082 * It returns the code point in utf8 for the value in *encp.
10083 * value: a code value in the source encoding
10084 * encp: a pointer to an Encode object
10086 * If the result from Encode is not a single character,
10087 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10090 S_reg_recode(pTHX_ const char value, SV **encp)
10093 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10094 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10095 const STRLEN newlen = SvCUR(sv);
10096 UV uv = UNICODE_REPLACEMENT;
10098 PERL_ARGS_ASSERT_REG_RECODE;
10102 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10105 if (!newlen || numlen != newlen) {
10106 uv = UNICODE_REPLACEMENT;
10112 PERL_STATIC_INLINE U8
10113 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10117 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10123 op = get_regex_charset(RExC_flags);
10124 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10125 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10126 been, so there is no hole */
10129 return op + EXACTF;
10132 PERL_STATIC_INLINE void
10133 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
10135 /* This knows the details about sizing an EXACTish node, setting flags for
10136 * it (by setting <*flagp>, and potentially populating it with a single
10139 * If <len> (the length in bytes) is non-zero, this function assumes that
10140 * the node has already been populated, and just does the sizing. In this
10141 * case <code_point> should be the final code point that has already been
10142 * placed into the node. This value will be ignored except that under some
10143 * circumstances <*flagp> is set based on it.
10145 * If <len> is zero, the function assumes that the node is to contain only
10146 * the single character given by <code_point> and calculates what <len>
10147 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10148 * additionally will populate the node's STRING with <code_point>, if <len>
10149 * is 0. In both cases <*flagp> is appropriately set
10151 * It knows that under FOLD, the Latin Sharp S and UTF characters above
10152 * 255, must be folded (the former only when the rules indicate it can
10155 bool len_passed_in = cBOOL(len != 0);
10156 U8 character[UTF8_MAXBYTES_CASE+1];
10158 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10160 if (! len_passed_in) {
10162 if (FOLD && (! LOC || code_point > 255)) {
10163 _to_uni_fold_flags(NATIVE_TO_UNI(code_point),
10166 FOLD_FLAGS_FULL | ((LOC)
10167 ? FOLD_FLAGS_LOCALE
10168 : (ASCII_FOLD_RESTRICTED)
10169 ? FOLD_FLAGS_NOMIX_ASCII
10173 uvchr_to_utf8( character, code_point);
10174 len = UTF8SKIP(character);
10178 || code_point != LATIN_SMALL_LETTER_SHARP_S
10179 || ASCII_FOLD_RESTRICTED
10180 || ! AT_LEAST_UNI_SEMANTICS)
10182 *character = (U8) code_point;
10187 *(character + 1) = 's';
10193 RExC_size += STR_SZ(len);
10196 RExC_emit += STR_SZ(len);
10197 STR_LEN(node) = len;
10198 if (! len_passed_in) {
10199 Copy((char *) character, STRING(node), len, char);
10203 *flagp |= HASWIDTH;
10205 /* A single character node is SIMPLE, except for the special-cased SHARP S
10207 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
10208 && (code_point != LATIN_SMALL_LETTER_SHARP_S
10209 || ! FOLD || ! DEPENDS_SEMANTICS))
10216 - regatom - the lowest level
10218 Try to identify anything special at the start of the pattern. If there
10219 is, then handle it as required. This may involve generating a single regop,
10220 such as for an assertion; or it may involve recursing, such as to
10221 handle a () structure.
10223 If the string doesn't start with something special then we gobble up
10224 as much literal text as we can.
10226 Once we have been able to handle whatever type of thing started the
10227 sequence, we return.
10229 Note: we have to be careful with escapes, as they can be both literal
10230 and special, and in the case of \10 and friends, context determines which.
10232 A summary of the code structure is:
10234 switch (first_byte) {
10235 cases for each special:
10236 handle this special;
10239 switch (2nd byte) {
10240 cases for each unambiguous special:
10241 handle this special;
10243 cases for each ambigous special/literal:
10245 if (special) handle here
10247 default: // unambiguously literal:
10250 default: // is a literal char
10253 create EXACTish node for literal;
10254 while (more input and node isn't full) {
10255 switch (input_byte) {
10256 cases for each special;
10257 make sure parse pointer is set so that the next call to
10258 regatom will see this special first
10259 goto loopdone; // EXACTish node terminated by prev. char
10261 append char to EXACTISH node;
10263 get next input byte;
10267 return the generated node;
10269 Specifically there are two separate switches for handling
10270 escape sequences, with the one for handling literal escapes requiring
10271 a dummy entry for all of the special escapes that are actually handled
10274 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
10276 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10278 Otherwise does not return NULL.
10282 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10285 regnode *ret = NULL;
10287 char *parse_start = RExC_parse;
10291 GET_RE_DEBUG_FLAGS_DECL;
10293 *flagp = WORST; /* Tentatively. */
10295 DEBUG_PARSE("atom");
10297 PERL_ARGS_ASSERT_REGATOM;
10300 switch ((U8)*RExC_parse) {
10302 RExC_seen_zerolen++;
10303 nextchar(pRExC_state);
10304 if (RExC_flags & RXf_PMf_MULTILINE)
10305 ret = reg_node(pRExC_state, MBOL);
10306 else if (RExC_flags & RXf_PMf_SINGLELINE)
10307 ret = reg_node(pRExC_state, SBOL);
10309 ret = reg_node(pRExC_state, BOL);
10310 Set_Node_Length(ret, 1); /* MJD */
10313 nextchar(pRExC_state);
10315 RExC_seen_zerolen++;
10316 if (RExC_flags & RXf_PMf_MULTILINE)
10317 ret = reg_node(pRExC_state, MEOL);
10318 else if (RExC_flags & RXf_PMf_SINGLELINE)
10319 ret = reg_node(pRExC_state, SEOL);
10321 ret = reg_node(pRExC_state, EOL);
10322 Set_Node_Length(ret, 1); /* MJD */
10325 nextchar(pRExC_state);
10326 if (RExC_flags & RXf_PMf_SINGLELINE)
10327 ret = reg_node(pRExC_state, SANY);
10329 ret = reg_node(pRExC_state, REG_ANY);
10330 *flagp |= HASWIDTH|SIMPLE;
10332 Set_Node_Length(ret, 1); /* MJD */
10336 char * const oregcomp_parse = ++RExC_parse;
10337 ret = regclass(pRExC_state, flagp,depth+1,
10338 FALSE, /* means parse the whole char class */
10339 TRUE, /* allow multi-char folds */
10340 FALSE, /* don't silence non-portable warnings. */
10342 if (*RExC_parse != ']') {
10343 RExC_parse = oregcomp_parse;
10344 vFAIL("Unmatched [");
10347 if (*flagp & RESTART_UTF8)
10349 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10352 nextchar(pRExC_state);
10353 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10357 nextchar(pRExC_state);
10358 ret = reg(pRExC_state, 2, &flags,depth+1);
10360 if (flags & TRYAGAIN) {
10361 if (RExC_parse == RExC_end) {
10362 /* Make parent create an empty node if needed. */
10363 *flagp |= TRYAGAIN;
10368 if (flags & RESTART_UTF8) {
10369 *flagp = RESTART_UTF8;
10372 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"", (UV) flags);
10374 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10378 if (flags & TRYAGAIN) {
10379 *flagp |= TRYAGAIN;
10382 vFAIL("Internal urp");
10383 /* Supposed to be caught earlier. */
10386 if (!regcurly(RExC_parse, FALSE)) {
10395 vFAIL("Quantifier follows nothing");
10400 This switch handles escape sequences that resolve to some kind
10401 of special regop and not to literal text. Escape sequnces that
10402 resolve to literal text are handled below in the switch marked
10405 Every entry in this switch *must* have a corresponding entry
10406 in the literal escape switch. However, the opposite is not
10407 required, as the default for this switch is to jump to the
10408 literal text handling code.
10410 switch ((U8)*++RExC_parse) {
10412 /* Special Escapes */
10414 RExC_seen_zerolen++;
10415 ret = reg_node(pRExC_state, SBOL);
10417 goto finish_meta_pat;
10419 ret = reg_node(pRExC_state, GPOS);
10420 RExC_seen |= REG_SEEN_GPOS;
10422 goto finish_meta_pat;
10424 RExC_seen_zerolen++;
10425 ret = reg_node(pRExC_state, KEEPS);
10427 /* XXX:dmq : disabling in-place substitution seems to
10428 * be necessary here to avoid cases of memory corruption, as
10429 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10431 RExC_seen |= REG_SEEN_LOOKBEHIND;
10432 goto finish_meta_pat;
10434 ret = reg_node(pRExC_state, SEOL);
10436 RExC_seen_zerolen++; /* Do not optimize RE away */
10437 goto finish_meta_pat;
10439 ret = reg_node(pRExC_state, EOS);
10441 RExC_seen_zerolen++; /* Do not optimize RE away */
10442 goto finish_meta_pat;
10444 ret = reg_node(pRExC_state, CANY);
10445 RExC_seen |= REG_SEEN_CANY;
10446 *flagp |= HASWIDTH|SIMPLE;
10447 goto finish_meta_pat;
10449 ret = reg_node(pRExC_state, CLUMP);
10450 *flagp |= HASWIDTH;
10451 goto finish_meta_pat;
10457 arg = ANYOF_WORDCHAR;
10461 RExC_seen_zerolen++;
10462 RExC_seen |= REG_SEEN_LOOKBEHIND;
10463 op = BOUND + get_regex_charset(RExC_flags);
10464 if (op > BOUNDA) { /* /aa is same as /a */
10467 ret = reg_node(pRExC_state, op);
10468 FLAGS(ret) = get_regex_charset(RExC_flags);
10470 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10471 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
10473 goto finish_meta_pat;
10475 RExC_seen_zerolen++;
10476 RExC_seen |= REG_SEEN_LOOKBEHIND;
10477 op = NBOUND + get_regex_charset(RExC_flags);
10478 if (op > NBOUNDA) { /* /aa is same as /a */
10481 ret = reg_node(pRExC_state, op);
10482 FLAGS(ret) = get_regex_charset(RExC_flags);
10484 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10485 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
10487 goto finish_meta_pat;
10497 ret = reg_node(pRExC_state, LNBREAK);
10498 *flagp |= HASWIDTH|SIMPLE;
10499 goto finish_meta_pat;
10507 goto join_posix_op_known;
10513 arg = ANYOF_VERTWS;
10515 goto join_posix_op_known;
10525 op = POSIXD + get_regex_charset(RExC_flags);
10526 if (op > POSIXA) { /* /aa is same as /a */
10530 join_posix_op_known:
10533 op += NPOSIXD - POSIXD;
10536 ret = reg_node(pRExC_state, op);
10538 FLAGS(ret) = namedclass_to_classnum(arg);
10541 *flagp |= HASWIDTH|SIMPLE;
10545 nextchar(pRExC_state);
10546 Set_Node_Length(ret, 2); /* MJD */
10552 char* parse_start = RExC_parse - 2;
10557 ret = regclass(pRExC_state, flagp,depth+1,
10558 TRUE, /* means just parse this element */
10559 FALSE, /* don't allow multi-char folds */
10560 FALSE, /* don't silence non-portable warnings.
10561 It would be a bug if these returned
10564 /* regclass() can only return RESTART_UTF8 if multi-char folds
10567 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10572 Set_Node_Offset(ret, parse_start + 2);
10573 Set_Node_Cur_Length(ret, parse_start);
10574 nextchar(pRExC_state);
10578 /* Handle \N and \N{NAME} with multiple code points here and not
10579 * below because it can be multicharacter. join_exact() will join
10580 * them up later on. Also this makes sure that things like
10581 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10582 * The options to the grok function call causes it to fail if the
10583 * sequence is just a single code point. We then go treat it as
10584 * just another character in the current EXACT node, and hence it
10585 * gets uniform treatment with all the other characters. The
10586 * special treatment for quantifiers is not needed for such single
10587 * character sequences */
10589 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
10590 FALSE /* not strict */ )) {
10591 if (*flagp & RESTART_UTF8)
10597 case 'k': /* Handle \k<NAME> and \k'NAME' */
10600 char ch= RExC_parse[1];
10601 if (ch != '<' && ch != '\'' && ch != '{') {
10603 vFAIL2("Sequence %.2s... not terminated",parse_start);
10605 /* this pretty much dupes the code for (?P=...) in reg(), if
10606 you change this make sure you change that */
10607 char* name_start = (RExC_parse += 2);
10609 SV *sv_dat = reg_scan_name(pRExC_state,
10610 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10611 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10612 if (RExC_parse == name_start || *RExC_parse != ch)
10613 vFAIL2("Sequence %.3s... not terminated",parse_start);
10616 num = add_data( pRExC_state, 1, "S" );
10617 RExC_rxi->data->data[num]=(void*)sv_dat;
10618 SvREFCNT_inc_simple_void(sv_dat);
10622 ret = reganode(pRExC_state,
10625 : (ASCII_FOLD_RESTRICTED)
10627 : (AT_LEAST_UNI_SEMANTICS)
10633 *flagp |= HASWIDTH;
10635 /* override incorrect value set in reganode MJD */
10636 Set_Node_Offset(ret, parse_start+1);
10637 Set_Node_Cur_Length(ret, parse_start);
10638 nextchar(pRExC_state);
10644 case '1': case '2': case '3': case '4':
10645 case '5': case '6': case '7': case '8': case '9':
10648 bool isg = *RExC_parse == 'g';
10653 if (*RExC_parse == '{') {
10657 if (*RExC_parse == '-') {
10661 if (hasbrace && !isDIGIT(*RExC_parse)) {
10662 if (isrel) RExC_parse--;
10664 goto parse_named_seq;
10666 num = atoi(RExC_parse);
10667 if (isg && num == 0) {
10668 if (*RExC_parse == '0') {
10669 vFAIL("Reference to invalid group 0");
10672 vFAIL("Unterminated \\g... pattern");
10676 num = RExC_npar - num;
10678 vFAIL("Reference to nonexistent or unclosed group");
10680 if (!isg && num > 9 && num >= RExC_npar && *RExC_parse != '8' && *RExC_parse != '9')
10681 /* Probably a character specified in octal, e.g. \35 */
10684 #ifdef RE_TRACK_PATTERN_OFFSETS
10685 char * const parse_start = RExC_parse - 1; /* MJD */
10687 while (isDIGIT(*RExC_parse))
10690 if (*RExC_parse != '}')
10691 vFAIL("Unterminated \\g{...} pattern");
10695 if (num > (I32)RExC_rx->nparens)
10696 vFAIL("Reference to nonexistent group");
10699 ret = reganode(pRExC_state,
10702 : (ASCII_FOLD_RESTRICTED)
10704 : (AT_LEAST_UNI_SEMANTICS)
10710 *flagp |= HASWIDTH;
10712 /* override incorrect value set in reganode MJD */
10713 Set_Node_Offset(ret, parse_start+1);
10714 Set_Node_Cur_Length(ret, parse_start);
10716 nextchar(pRExC_state);
10721 if (RExC_parse >= RExC_end)
10722 FAIL("Trailing \\");
10725 /* Do not generate "unrecognized" warnings here, we fall
10726 back into the quick-grab loop below */
10733 if (RExC_flags & RXf_PMf_EXTENDED) {
10734 if ( reg_skipcomment( pRExC_state ) )
10741 parse_start = RExC_parse - 1;
10750 #define MAX_NODE_STRING_SIZE 127
10751 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10753 U8 upper_parse = MAX_NODE_STRING_SIZE;
10756 bool next_is_quantifier;
10757 char * oldp = NULL;
10759 /* If a folding node contains only code points that don't
10760 * participate in folds, it can be changed into an EXACT node,
10761 * which allows the optimizer more things to look for */
10765 node_type = compute_EXACTish(pRExC_state);
10766 ret = reg_node(pRExC_state, node_type);
10768 /* In pass1, folded, we use a temporary buffer instead of the
10769 * actual node, as the node doesn't exist yet */
10770 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10776 /* We do the EXACTFish to EXACT node only if folding, and not if in
10777 * locale, as whether a character folds or not isn't known until
10779 maybe_exact = FOLD && ! LOC;
10781 /* XXX The node can hold up to 255 bytes, yet this only goes to
10782 * 127. I (khw) do not know why. Keeping it somewhat less than
10783 * 255 allows us to not have to worry about overflow due to
10784 * converting to utf8 and fold expansion, but that value is
10785 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10786 * split up by this limit into a single one using the real max of
10787 * 255. Even at 127, this breaks under rare circumstances. If
10788 * folding, we do not want to split a node at a character that is a
10789 * non-final in a multi-char fold, as an input string could just
10790 * happen to want to match across the node boundary. The join
10791 * would solve that problem if the join actually happens. But a
10792 * series of more than two nodes in a row each of 127 would cause
10793 * the first join to succeed to get to 254, but then there wouldn't
10794 * be room for the next one, which could at be one of those split
10795 * multi-char folds. I don't know of any fool-proof solution. One
10796 * could back off to end with only a code point that isn't such a
10797 * non-final, but it is possible for there not to be any in the
10799 for (p = RExC_parse - 1;
10800 len < upper_parse && p < RExC_end;
10805 if (RExC_flags & RXf_PMf_EXTENDED)
10806 p = regwhite( pRExC_state, p );
10817 /* Literal Escapes Switch
10819 This switch is meant to handle escape sequences that
10820 resolve to a literal character.
10822 Every escape sequence that represents something
10823 else, like an assertion or a char class, is handled
10824 in the switch marked 'Special Escapes' above in this
10825 routine, but also has an entry here as anything that
10826 isn't explicitly mentioned here will be treated as
10827 an unescaped equivalent literal.
10830 switch ((U8)*++p) {
10831 /* These are all the special escapes. */
10832 case 'A': /* Start assertion */
10833 case 'b': case 'B': /* Word-boundary assertion*/
10834 case 'C': /* Single char !DANGEROUS! */
10835 case 'd': case 'D': /* digit class */
10836 case 'g': case 'G': /* generic-backref, pos assertion */
10837 case 'h': case 'H': /* HORIZWS */
10838 case 'k': case 'K': /* named backref, keep marker */
10839 case 'p': case 'P': /* Unicode property */
10840 case 'R': /* LNBREAK */
10841 case 's': case 'S': /* space class */
10842 case 'v': case 'V': /* VERTWS */
10843 case 'w': case 'W': /* word class */
10844 case 'X': /* eXtended Unicode "combining character sequence" */
10845 case 'z': case 'Z': /* End of line/string assertion */
10849 /* Anything after here is an escape that resolves to a
10850 literal. (Except digits, which may or may not)
10856 case 'N': /* Handle a single-code point named character. */
10857 /* The options cause it to fail if a multiple code
10858 * point sequence. Handle those in the switch() above
10860 RExC_parse = p + 1;
10861 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10862 flagp, depth, FALSE,
10863 FALSE /* not strict */ ))
10865 if (*flagp & RESTART_UTF8)
10866 FAIL("panic: grok_bslash_N set RESTART_UTF8");
10867 RExC_parse = p = oldp;
10871 if (ender > 0xff) {
10888 ender = ASCII_TO_NATIVE('\033');
10892 ender = ASCII_TO_NATIVE('\007');
10898 const char* error_msg;
10900 bool valid = grok_bslash_o(&p,
10903 TRUE, /* out warnings */
10904 FALSE, /* not strict */
10905 TRUE, /* Output warnings
10910 RExC_parse = p; /* going to die anyway; point
10911 to exact spot of failure */
10915 if (PL_encoding && ender < 0x100) {
10916 goto recode_encoding;
10918 if (ender > 0xff) {
10925 UV result = UV_MAX; /* initialize to erroneous
10927 const char* error_msg;
10929 bool valid = grok_bslash_x(&p,
10932 TRUE, /* out warnings */
10933 FALSE, /* not strict */
10934 TRUE, /* Output warnings
10939 RExC_parse = p; /* going to die anyway; point
10940 to exact spot of failure */
10945 if (PL_encoding && ender < 0x100) {
10946 goto recode_encoding;
10948 if (ender > 0xff) {
10955 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10957 case '8': case '9': /* must be a backreference */
10960 case '1': case '2': case '3':case '4':
10961 case '5': case '6': case '7':
10962 /* When we parse backslash escapes there is ambiguity between
10963 * backreferences and octal escapes. Any escape from \1 - \9 is
10964 * a backreference, any multi-digit escape which does not start with
10965 * 0 and which when evaluated as decimal could refer to an already
10966 * parsed capture buffer is a backslash. Anything else is octal.
10968 * Note this implies that \118 could be interpreted as 118 OR as
10969 * "\11" . "8" depending on whether there were 118 capture buffers
10970 * defined already in the pattern.
10972 if ( !isDIGIT(p[1]) || atoi(p) <= RExC_npar )
10973 { /* Not to be treated as an octal constant, go
10980 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10982 ender = grok_oct(p, &numlen, &flags, NULL);
10983 if (ender > 0xff) {
10987 if (SIZE_ONLY /* like \08, \178 */
10990 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
10992 reg_warn_non_literal_string(
10994 form_short_octal_warning(p, numlen));
10997 if (PL_encoding && ender < 0x100)
10998 goto recode_encoding;
11001 if (! RExC_override_recoding) {
11002 SV* enc = PL_encoding;
11003 ender = reg_recode((const char)(U8)ender, &enc);
11004 if (!enc && SIZE_ONLY)
11005 ckWARNreg(p, "Invalid escape in the specified encoding");
11011 FAIL("Trailing \\");
11014 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
11015 /* Include any { following the alpha to emphasize
11016 * that it could be part of an escape at some point
11018 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
11019 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
11021 goto normal_default;
11022 } /* End of switch on '\' */
11024 default: /* A literal character */
11027 && RExC_flags & RXf_PMf_EXTENDED
11028 && ckWARN_d(WARN_DEPRECATED)
11029 && is_PATWS_non_low(p, UTF))
11031 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
11032 "Escape literal pattern white space under /x");
11036 if (UTF8_IS_START(*p) && UTF) {
11038 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
11039 &numlen, UTF8_ALLOW_DEFAULT);
11045 } /* End of switch on the literal */
11047 /* Here, have looked at the literal character and <ender>
11048 * contains its ordinal, <p> points to the character after it
11051 if ( RExC_flags & RXf_PMf_EXTENDED)
11052 p = regwhite( pRExC_state, p );
11054 /* If the next thing is a quantifier, it applies to this
11055 * character only, which means that this character has to be in
11056 * its own node and can't just be appended to the string in an
11057 * existing node, so if there are already other characters in
11058 * the node, close the node with just them, and set up to do
11059 * this character again next time through, when it will be the
11060 * only thing in its new node */
11061 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
11069 const STRLEN unilen = reguni(pRExC_state, ender, s);
11075 /* The loop increments <len> each time, as all but this
11076 * path (and one other) through it add a single byte to
11077 * the EXACTish node. But this one has changed len to
11078 * be the correct final value, so subtract one to
11079 * cancel out the increment that follows */
11083 REGC((char)ender, s++);
11088 /* See comments for join_exact() as to why we fold this
11089 * non-UTF at compile time */
11090 || (node_type == EXACTFU
11091 && ender == LATIN_SMALL_LETTER_SHARP_S)))
11093 *(s++) = (char) ender;
11094 maybe_exact &= ! IS_IN_SOME_FOLD_L1(ender);
11098 /* Prime the casefolded buffer. Locale rules, which apply
11099 * only to code points < 256, aren't known until execution,
11100 * so for them, just output the original character using
11101 * utf8. If we start to fold non-UTF patterns, be sure to
11102 * update join_exact() */
11103 if (LOC && ender < 256) {
11104 if (UNI_IS_INVARIANT(ender)) {
11108 *s = UTF8_TWO_BYTE_HI(ender);
11109 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
11114 UV folded = _to_uni_fold_flags(
11119 | ((LOC) ? FOLD_FLAGS_LOCALE
11120 : (ASCII_FOLD_RESTRICTED)
11121 ? FOLD_FLAGS_NOMIX_ASCII
11125 /* If this node only contains non-folding code points
11126 * so far, see if this new one is also non-folding */
11128 if (folded != ender) {
11129 maybe_exact = FALSE;
11132 /* Here the fold is the original; we have
11133 * to check further to see if anything
11135 if (! PL_utf8_foldable) {
11136 SV* swash = swash_init("utf8",
11138 &PL_sv_undef, 1, 0);
11140 _get_swash_invlist(swash);
11141 SvREFCNT_dec_NN(swash);
11143 if (_invlist_contains_cp(PL_utf8_foldable,
11146 maybe_exact = FALSE;
11154 /* The loop increments <len> each time, as all but this
11155 * path (and one other) through it add a single byte to the
11156 * EXACTish node. But this one has changed len to be the
11157 * correct final value, so subtract one to cancel out the
11158 * increment that follows */
11159 len += foldlen - 1;
11162 if (next_is_quantifier) {
11164 /* Here, the next input is a quantifier, and to get here,
11165 * the current character is the only one in the node.
11166 * Also, here <len> doesn't include the final byte for this
11172 } /* End of loop through literal characters */
11174 /* Here we have either exhausted the input or ran out of room in
11175 * the node. (If we encountered a character that can't be in the
11176 * node, transfer is made directly to <loopdone>, and so we
11177 * wouldn't have fallen off the end of the loop.) In the latter
11178 * case, we artificially have to split the node into two, because
11179 * we just don't have enough space to hold everything. This
11180 * creates a problem if the final character participates in a
11181 * multi-character fold in the non-final position, as a match that
11182 * should have occurred won't, due to the way nodes are matched,
11183 * and our artificial boundary. So back off until we find a non-
11184 * problematic character -- one that isn't at the beginning or
11185 * middle of such a fold. (Either it doesn't participate in any
11186 * folds, or appears only in the final position of all the folds it
11187 * does participate in.) A better solution with far fewer false
11188 * positives, and that would fill the nodes more completely, would
11189 * be to actually have available all the multi-character folds to
11190 * test against, and to back-off only far enough to be sure that
11191 * this node isn't ending with a partial one. <upper_parse> is set
11192 * further below (if we need to reparse the node) to include just
11193 * up through that final non-problematic character that this code
11194 * identifies, so when it is set to less than the full node, we can
11195 * skip the rest of this */
11196 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
11198 const STRLEN full_len = len;
11200 assert(len >= MAX_NODE_STRING_SIZE);
11202 /* Here, <s> points to the final byte of the final character.
11203 * Look backwards through the string until find a non-
11204 * problematic character */
11208 /* These two have no multi-char folds to non-UTF characters
11210 if (ASCII_FOLD_RESTRICTED || LOC) {
11214 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
11218 if (! PL_NonL1NonFinalFold) {
11219 PL_NonL1NonFinalFold = _new_invlist_C_array(
11220 NonL1_Perl_Non_Final_Folds_invlist);
11223 /* Point to the first byte of the final character */
11224 s = (char *) utf8_hop((U8 *) s, -1);
11226 while (s >= s0) { /* Search backwards until find
11227 non-problematic char */
11228 if (UTF8_IS_INVARIANT(*s)) {
11230 /* There are no ascii characters that participate
11231 * in multi-char folds under /aa. In EBCDIC, the
11232 * non-ascii invariants are all control characters,
11233 * so don't ever participate in any folds. */
11234 if (ASCII_FOLD_RESTRICTED
11235 || ! IS_NON_FINAL_FOLD(*s))
11240 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
11242 /* No Latin1 characters participate in multi-char
11243 * folds under /l */
11245 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
11251 else if (! _invlist_contains_cp(
11252 PL_NonL1NonFinalFold,
11253 valid_utf8_to_uvchr((U8 *) s, NULL)))
11258 /* Here, the current character is problematic in that
11259 * it does occur in the non-final position of some
11260 * fold, so try the character before it, but have to
11261 * special case the very first byte in the string, so
11262 * we don't read outside the string */
11263 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
11264 } /* End of loop backwards through the string */
11266 /* If there were only problematic characters in the string,
11267 * <s> will point to before s0, in which case the length
11268 * should be 0, otherwise include the length of the
11269 * non-problematic character just found */
11270 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
11273 /* Here, have found the final character, if any, that is
11274 * non-problematic as far as ending the node without splitting
11275 * it across a potential multi-char fold. <len> contains the
11276 * number of bytes in the node up-to and including that
11277 * character, or is 0 if there is no such character, meaning
11278 * the whole node contains only problematic characters. In
11279 * this case, give up and just take the node as-is. We can't
11285 /* Here, the node does contain some characters that aren't
11286 * problematic. If one such is the final character in the
11287 * node, we are done */
11288 if (len == full_len) {
11291 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
11293 /* If the final character is problematic, but the
11294 * penultimate is not, back-off that last character to
11295 * later start a new node with it */
11300 /* Here, the final non-problematic character is earlier
11301 * in the input than the penultimate character. What we do
11302 * is reparse from the beginning, going up only as far as
11303 * this final ok one, thus guaranteeing that the node ends
11304 * in an acceptable character. The reason we reparse is
11305 * that we know how far in the character is, but we don't
11306 * know how to correlate its position with the input parse.
11307 * An alternate implementation would be to build that
11308 * correlation as we go along during the original parse,
11309 * but that would entail extra work for every node, whereas
11310 * this code gets executed only when the string is too
11311 * large for the node, and the final two characters are
11312 * problematic, an infrequent occurrence. Yet another
11313 * possible strategy would be to save the tail of the
11314 * string, and the next time regatom is called, initialize
11315 * with that. The problem with this is that unless you
11316 * back off one more character, you won't be guaranteed
11317 * regatom will get called again, unless regbranch,
11318 * regpiece ... are also changed. If you do back off that
11319 * extra character, so that there is input guaranteed to
11320 * force calling regatom, you can't handle the case where
11321 * just the first character in the node is acceptable. I
11322 * (khw) decided to try this method which doesn't have that
11323 * pitfall; if performance issues are found, we can do a
11324 * combination of the current approach plus that one */
11330 } /* End of verifying node ends with an appropriate char */
11332 loopdone: /* Jumped to when encounters something that shouldn't be in
11335 /* I (khw) don't know if you can get here with zero length, but the
11336 * old code handled this situation by creating a zero-length EXACT
11337 * node. Might as well be NOTHING instead */
11343 /* If 'maybe_exact' is still set here, means there are no
11344 * code points in the node that participate in folds */
11345 if (FOLD && maybe_exact) {
11348 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11351 RExC_parse = p - 1;
11352 Set_Node_Cur_Length(ret, parse_start);
11353 nextchar(pRExC_state);
11355 /* len is STRLEN which is unsigned, need to copy to signed */
11358 vFAIL("Internal disaster");
11361 } /* End of label 'defchar:' */
11363 } /* End of giant switch on input character */
11369 S_regwhite( RExC_state_t *pRExC_state, char *p )
11371 const char *e = RExC_end;
11373 PERL_ARGS_ASSERT_REGWHITE;
11378 else if (*p == '#') {
11381 if (*p++ == '\n') {
11387 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11396 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
11398 /* Returns the next non-pattern-white space, non-comment character (the
11399 * latter only if 'recognize_comment is true) in the string p, which is
11400 * ended by RExC_end. If there is no line break ending a comment,
11401 * RExC_seen has added the REG_SEEN_RUN_ON_COMMENT flag; */
11402 const char *e = RExC_end;
11404 PERL_ARGS_ASSERT_REGPATWS;
11408 if ((len = is_PATWS_safe(p, e, UTF))) {
11411 else if (recognize_comment && *p == '#') {
11415 if (is_LNBREAK_safe(p, e, UTF)) {
11421 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11429 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11430 Character classes ([:foo:]) can also be negated ([:^foo:]).
11431 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11432 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11433 but trigger failures because they are currently unimplemented. */
11435 #define POSIXCC_DONE(c) ((c) == ':')
11436 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11437 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11439 PERL_STATIC_INLINE I32
11440 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
11443 I32 namedclass = OOB_NAMEDCLASS;
11445 PERL_ARGS_ASSERT_REGPPOSIXCC;
11447 if (value == '[' && RExC_parse + 1 < RExC_end &&
11448 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11449 POSIXCC(UCHARAT(RExC_parse)))
11451 const char c = UCHARAT(RExC_parse);
11452 char* const s = RExC_parse++;
11454 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11456 if (RExC_parse == RExC_end) {
11459 /* Try to give a better location for the error (than the end of
11460 * the string) by looking for the matching ']' */
11462 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11465 vFAIL2("Unmatched '%c' in POSIX class", c);
11467 /* Grandfather lone [:, [=, [. */
11471 const char* const t = RExC_parse++; /* skip over the c */
11474 if (UCHARAT(RExC_parse) == ']') {
11475 const char *posixcc = s + 1;
11476 RExC_parse++; /* skip over the ending ] */
11479 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11480 const I32 skip = t - posixcc;
11482 /* Initially switch on the length of the name. */
11485 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
11486 this is the Perl \w
11488 namedclass = ANYOF_WORDCHAR;
11491 /* Names all of length 5. */
11492 /* alnum alpha ascii blank cntrl digit graph lower
11493 print punct space upper */
11494 /* Offset 4 gives the best switch position. */
11495 switch (posixcc[4]) {
11497 if (memEQ(posixcc, "alph", 4)) /* alpha */
11498 namedclass = ANYOF_ALPHA;
11501 if (memEQ(posixcc, "spac", 4)) /* space */
11502 namedclass = ANYOF_PSXSPC;
11505 if (memEQ(posixcc, "grap", 4)) /* graph */
11506 namedclass = ANYOF_GRAPH;
11509 if (memEQ(posixcc, "asci", 4)) /* ascii */
11510 namedclass = ANYOF_ASCII;
11513 if (memEQ(posixcc, "blan", 4)) /* blank */
11514 namedclass = ANYOF_BLANK;
11517 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11518 namedclass = ANYOF_CNTRL;
11521 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11522 namedclass = ANYOF_ALPHANUMERIC;
11525 if (memEQ(posixcc, "lowe", 4)) /* lower */
11526 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
11527 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11528 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
11531 if (memEQ(posixcc, "digi", 4)) /* digit */
11532 namedclass = ANYOF_DIGIT;
11533 else if (memEQ(posixcc, "prin", 4)) /* print */
11534 namedclass = ANYOF_PRINT;
11535 else if (memEQ(posixcc, "punc", 4)) /* punct */
11536 namedclass = ANYOF_PUNCT;
11541 if (memEQ(posixcc, "xdigit", 6))
11542 namedclass = ANYOF_XDIGIT;
11546 if (namedclass == OOB_NAMEDCLASS)
11547 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11550 /* The #defines are structured so each complement is +1 to
11551 * the normal one */
11555 assert (posixcc[skip] == ':');
11556 assert (posixcc[skip+1] == ']');
11557 } else if (!SIZE_ONLY) {
11558 /* [[=foo=]] and [[.foo.]] are still future. */
11560 /* adjust RExC_parse so the warning shows after
11561 the class closes */
11562 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11564 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11567 /* Maternal grandfather:
11568 * "[:" ending in ":" but not in ":]" */
11570 vFAIL("Unmatched '[' in POSIX class");
11573 /* Grandfather lone [:, [=, [. */
11583 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
11585 /* This applies some heuristics at the current parse position (which should
11586 * be at a '[') to see if what follows might be intended to be a [:posix:]
11587 * class. It returns true if it really is a posix class, of course, but it
11588 * also can return true if it thinks that what was intended was a posix
11589 * class that didn't quite make it.
11591 * It will return true for
11593 * [:alphanumerics] (as long as the ] isn't followed immediately by a
11594 * ')' indicating the end of the (?[
11595 * [:any garbage including %^&$ punctuation:]
11597 * This is designed to be called only from S_handle_regex_sets; it could be
11598 * easily adapted to be called from the spot at the beginning of regclass()
11599 * that checks to see in a normal bracketed class if the surrounding []
11600 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
11601 * change long-standing behavior, so I (khw) didn't do that */
11602 char* p = RExC_parse + 1;
11603 char first_char = *p;
11605 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
11607 assert(*(p - 1) == '[');
11609 if (! POSIXCC(first_char)) {
11614 while (p < RExC_end && isWORDCHAR(*p)) p++;
11616 if (p >= RExC_end) {
11620 if (p - RExC_parse > 2 /* Got at least 1 word character */
11621 && (*p == first_char
11622 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
11627 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
11630 && p - RExC_parse > 2 /* [:] evaluates to colon;
11631 [::] is a bad posix class. */
11632 && first_char == *(p - 1));
11636 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist, I32 *flagp, U32 depth,
11637 char * const oregcomp_parse)
11639 /* Handle the (?[...]) construct to do set operations */
11642 UV start, end; /* End points of code point ranges */
11644 char *save_end, *save_parse;
11649 const bool save_fold = FOLD;
11651 GET_RE_DEBUG_FLAGS_DECL;
11653 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
11656 vFAIL("(?[...]) not valid in locale");
11658 RExC_uni_semantics = 1;
11660 /* This will return only an ANYOF regnode, or (unlikely) something smaller
11661 * (such as EXACT). Thus we can skip most everything if just sizing. We
11662 * call regclass to handle '[]' so as to not have to reinvent its parsing
11663 * rules here (throwing away the size it computes each time). And, we exit
11664 * upon an unescaped ']' that isn't one ending a regclass. To do both
11665 * these things, we need to realize that something preceded by a backslash
11666 * is escaped, so we have to keep track of backslashes */
11668 UV depth = 0; /* how many nested (?[...]) constructs */
11670 Perl_ck_warner_d(aTHX_
11671 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
11672 "The regex_sets feature is experimental" REPORT_LOCATION,
11673 (int) (RExC_parse - RExC_precomp) , RExC_precomp, RExC_parse);
11675 while (RExC_parse < RExC_end) {
11676 SV* current = NULL;
11677 RExC_parse = regpatws(pRExC_state, RExC_parse,
11678 TRUE); /* means recognize comments */
11679 switch (*RExC_parse) {
11681 if (RExC_parse[1] == '[') depth++, RExC_parse++;
11686 /* Skip the next byte (which could cause us to end up in
11687 * the middle of a UTF-8 character, but since none of those
11688 * are confusable with anything we currently handle in this
11689 * switch (invariants all), it's safe. We'll just hit the
11690 * default: case next time and keep on incrementing until
11691 * we find one of the invariants we do handle. */
11696 /* If this looks like it is a [:posix:] class, leave the
11697 * parse pointer at the '[' to fool regclass() into
11698 * thinking it is part of a '[[:posix:]]'. That function
11699 * will use strict checking to force a syntax error if it
11700 * doesn't work out to a legitimate class */
11701 bool is_posix_class
11702 = could_it_be_a_POSIX_class(pRExC_state);
11703 if (! is_posix_class) {
11707 /* regclass() can only return RESTART_UTF8 if multi-char
11708 folds are allowed. */
11709 if (!regclass(pRExC_state, flagp,depth+1,
11710 is_posix_class, /* parse the whole char
11711 class only if not a
11713 FALSE, /* don't allow multi-char folds */
11714 TRUE, /* silence non-portable warnings. */
11716 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11719 /* function call leaves parse pointing to the ']', except
11720 * if we faked it */
11721 if (is_posix_class) {
11725 SvREFCNT_dec(current); /* In case it returned something */
11730 if (depth--) break;
11732 if (RExC_parse < RExC_end
11733 && *RExC_parse == ')')
11735 node = reganode(pRExC_state, ANYOF, 0);
11736 RExC_size += ANYOF_SKIP;
11737 nextchar(pRExC_state);
11738 Set_Node_Length(node,
11739 RExC_parse - oregcomp_parse + 1); /* MJD */
11748 FAIL("Syntax error in (?[...])");
11751 /* Pass 2 only after this. Everything in this construct is a
11752 * metacharacter. Operands begin with either a '\' (for an escape
11753 * sequence), or a '[' for a bracketed character class. Any other
11754 * character should be an operator, or parenthesis for grouping. Both
11755 * types of operands are handled by calling regclass() to parse them. It
11756 * is called with a parameter to indicate to return the computed inversion
11757 * list. The parsing here is implemented via a stack. Each entry on the
11758 * stack is a single character representing one of the operators, or the
11759 * '('; or else a pointer to an operand inversion list. */
11761 #define IS_OPERAND(a) (! SvIOK(a))
11763 /* The stack starts empty. It is a syntax error if the first thing parsed
11764 * is a binary operator; everything else is pushed on the stack. When an
11765 * operand is parsed, the top of the stack is examined. If it is a binary
11766 * operator, the item before it should be an operand, and both are replaced
11767 * by the result of doing that operation on the new operand and the one on
11768 * the stack. Thus a sequence of binary operands is reduced to a single
11769 * one before the next one is parsed.
11771 * A unary operator may immediately follow a binary in the input, for
11774 * When an operand is parsed and the top of the stack is a unary operator,
11775 * the operation is performed, and then the stack is rechecked to see if
11776 * this new operand is part of a binary operation; if so, it is handled as
11779 * A '(' is simply pushed on the stack; it is valid only if the stack is
11780 * empty, or the top element of the stack is an operator or another '('
11781 * (for which the parenthesized expression will become an operand). By the
11782 * time the corresponding ')' is parsed everything in between should have
11783 * been parsed and evaluated to a single operand (or else is a syntax
11784 * error), and is handled as a regular operand */
11786 sv_2mortal((SV *)(stack = newAV()));
11788 while (RExC_parse < RExC_end) {
11789 I32 top_index = av_tindex(stack);
11791 SV* current = NULL;
11793 /* Skip white space */
11794 RExC_parse = regpatws(pRExC_state, RExC_parse,
11795 TRUE); /* means recognize comments */
11796 if (RExC_parse >= RExC_end) {
11797 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
11799 if ((curchar = UCHARAT(RExC_parse)) == ']') {
11806 if (av_tindex(stack) >= 0 /* This makes sure that we can
11807 safely subtract 1 from
11808 RExC_parse in the next clause.
11809 If we have something on the
11810 stack, we have parsed something
11812 && UCHARAT(RExC_parse - 1) == '('
11813 && RExC_parse < RExC_end)
11815 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
11816 * This happens when we have some thing like
11818 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
11820 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
11822 * Here we would be handling the interpolated
11823 * '$thai_or_lao'. We handle this by a recursive call to
11824 * ourselves which returns the inversion list the
11825 * interpolated expression evaluates to. We use the flags
11826 * from the interpolated pattern. */
11827 U32 save_flags = RExC_flags;
11828 const char * const save_parse = ++RExC_parse;
11830 parse_lparen_question_flags(pRExC_state);
11832 if (RExC_parse == save_parse /* Makes sure there was at
11833 least one flag (or this
11834 embedding wasn't compiled)
11836 || RExC_parse >= RExC_end - 4
11837 || UCHARAT(RExC_parse) != ':'
11838 || UCHARAT(++RExC_parse) != '('
11839 || UCHARAT(++RExC_parse) != '?'
11840 || UCHARAT(++RExC_parse) != '[')
11843 /* In combination with the above, this moves the
11844 * pointer to the point just after the first erroneous
11845 * character (or if there are no flags, to where they
11846 * should have been) */
11847 if (RExC_parse >= RExC_end - 4) {
11848 RExC_parse = RExC_end;
11850 else if (RExC_parse != save_parse) {
11851 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11853 vFAIL("Expecting '(?flags:(?[...'");
11856 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
11857 depth+1, oregcomp_parse);
11859 /* Here, 'current' contains the embedded expression's
11860 * inversion list, and RExC_parse points to the trailing
11861 * ']'; the next character should be the ')' which will be
11862 * paired with the '(' that has been put on the stack, so
11863 * the whole embedded expression reduces to '(operand)' */
11866 RExC_flags = save_flags;
11867 goto handle_operand;
11872 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11873 vFAIL("Unexpected character");
11876 /* regclass() can only return RESTART_UTF8 if multi-char
11877 folds are allowed. */
11878 if (!regclass(pRExC_state, flagp,depth+1,
11879 TRUE, /* means parse just the next thing */
11880 FALSE, /* don't allow multi-char folds */
11881 FALSE, /* don't silence non-portable warnings. */
11883 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11885 /* regclass() will return with parsing just the \ sequence,
11886 * leaving the parse pointer at the next thing to parse */
11888 goto handle_operand;
11890 case '[': /* Is a bracketed character class */
11892 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
11894 if (! is_posix_class) {
11898 /* regclass() can only return RESTART_UTF8 if multi-char
11899 folds are allowed. */
11900 if(!regclass(pRExC_state, flagp,depth+1,
11901 is_posix_class, /* parse the whole char class
11902 only if not a posix class */
11903 FALSE, /* don't allow multi-char folds */
11904 FALSE, /* don't silence non-portable warnings. */
11906 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11908 /* function call leaves parse pointing to the ']', except if we
11910 if (is_posix_class) {
11914 goto handle_operand;
11923 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
11924 || ! IS_OPERAND(*top_ptr))
11927 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
11929 av_push(stack, newSVuv(curchar));
11933 av_push(stack, newSVuv(curchar));
11937 if (top_index >= 0) {
11938 top_ptr = av_fetch(stack, top_index, FALSE);
11940 if (IS_OPERAND(*top_ptr)) {
11942 vFAIL("Unexpected '(' with no preceding operator");
11945 av_push(stack, newSVuv(curchar));
11952 || ! (current = av_pop(stack))
11953 || ! IS_OPERAND(current)
11954 || ! (lparen = av_pop(stack))
11955 || IS_OPERAND(lparen)
11956 || SvUV(lparen) != '(')
11958 SvREFCNT_dec(current);
11960 vFAIL("Unexpected ')'");
11963 SvREFCNT_dec_NN(lparen);
11970 /* Here, we have an operand to process, in 'current' */
11972 if (top_index < 0) { /* Just push if stack is empty */
11973 av_push(stack, current);
11976 SV* top = av_pop(stack);
11978 char current_operator;
11980 if (IS_OPERAND(top)) {
11981 SvREFCNT_dec_NN(top);
11982 SvREFCNT_dec_NN(current);
11983 vFAIL("Operand with no preceding operator");
11985 current_operator = (char) SvUV(top);
11986 switch (current_operator) {
11987 case '(': /* Push the '(' back on followed by the new
11989 av_push(stack, top);
11990 av_push(stack, current);
11991 SvREFCNT_inc(top); /* Counters the '_dec' done
11992 just after the 'break', so
11993 it doesn't get wrongly freed
11998 _invlist_invert(current);
12000 /* Unlike binary operators, the top of the stack,
12001 * now that this unary one has been popped off, may
12002 * legally be an operator, and we now have operand
12005 SvREFCNT_dec_NN(top);
12006 goto handle_operand;
12009 prev = av_pop(stack);
12010 _invlist_intersection(prev,
12013 av_push(stack, current);
12018 prev = av_pop(stack);
12019 _invlist_union(prev, current, ¤t);
12020 av_push(stack, current);
12024 prev = av_pop(stack);;
12025 _invlist_subtract(prev, current, ¤t);
12026 av_push(stack, current);
12029 case '^': /* The union minus the intersection */
12035 prev = av_pop(stack);
12036 _invlist_union(prev, current, &u);
12037 _invlist_intersection(prev, current, &i);
12038 /* _invlist_subtract will overwrite current
12039 without freeing what it already contains */
12041 _invlist_subtract(u, i, ¤t);
12042 av_push(stack, current);
12043 SvREFCNT_dec_NN(i);
12044 SvREFCNT_dec_NN(u);
12045 SvREFCNT_dec_NN(element);
12050 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
12052 SvREFCNT_dec_NN(top);
12053 SvREFCNT_dec(prev);
12057 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12060 if (av_tindex(stack) < 0 /* Was empty */
12061 || ((final = av_pop(stack)) == NULL)
12062 || ! IS_OPERAND(final)
12063 || av_tindex(stack) >= 0) /* More left on stack */
12065 vFAIL("Incomplete expression within '(?[ ])'");
12068 /* Here, 'final' is the resultant inversion list from evaluating the
12069 * expression. Return it if so requested */
12070 if (return_invlist) {
12071 *return_invlist = final;
12075 /* Otherwise generate a resultant node, based on 'final'. regclass() is
12076 * expecting a string of ranges and individual code points */
12077 invlist_iterinit(final);
12078 result_string = newSVpvs("");
12079 while (invlist_iternext(final, &start, &end)) {
12080 if (start == end) {
12081 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
12084 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
12089 save_parse = RExC_parse;
12090 RExC_parse = SvPV(result_string, len);
12091 save_end = RExC_end;
12092 RExC_end = RExC_parse + len;
12094 /* We turn off folding around the call, as the class we have constructed
12095 * already has all folding taken into consideration, and we don't want
12096 * regclass() to add to that */
12097 RExC_flags &= ~RXf_PMf_FOLD;
12098 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
12100 node = regclass(pRExC_state, flagp,depth+1,
12101 FALSE, /* means parse the whole char class */
12102 FALSE, /* don't allow multi-char folds */
12103 TRUE, /* silence non-portable warnings. The above may very
12104 well have generated non-portable code points, but
12105 they're valid on this machine */
12108 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
12111 RExC_flags |= RXf_PMf_FOLD;
12113 RExC_parse = save_parse + 1;
12114 RExC_end = save_end;
12115 SvREFCNT_dec_NN(final);
12116 SvREFCNT_dec_NN(result_string);
12118 nextchar(pRExC_state);
12119 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
12124 /* The names of properties whose definitions are not known at compile time are
12125 * stored in this SV, after a constant heading. So if the length has been
12126 * changed since initialization, then there is a run-time definition. */
12127 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
12130 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
12131 const bool stop_at_1, /* Just parse the next thing, don't
12132 look for a full character class */
12133 bool allow_multi_folds,
12134 const bool silence_non_portable, /* Don't output warnings
12137 SV** ret_invlist) /* Return an inversion list, not a node */
12139 /* parse a bracketed class specification. Most of these will produce an
12140 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
12141 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
12142 * under /i with multi-character folds: it will be rewritten following the
12143 * paradigm of this example, where the <multi-fold>s are characters which
12144 * fold to multiple character sequences:
12145 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
12146 * gets effectively rewritten as:
12147 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
12148 * reg() gets called (recursively) on the rewritten version, and this
12149 * function will return what it constructs. (Actually the <multi-fold>s
12150 * aren't physically removed from the [abcdefghi], it's just that they are
12151 * ignored in the recursion by means of a flag:
12152 * <RExC_in_multi_char_class>.)
12154 * ANYOF nodes contain a bit map for the first 256 characters, with the
12155 * corresponding bit set if that character is in the list. For characters
12156 * above 255, a range list or swash is used. There are extra bits for \w,
12157 * etc. in locale ANYOFs, as what these match is not determinable at
12160 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
12161 * to be restarted. This can only happen if ret_invlist is non-NULL.
12165 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
12167 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
12170 IV namedclass = OOB_NAMEDCLASS;
12171 char *rangebegin = NULL;
12172 bool need_class = 0;
12174 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
12175 than just initialized. */
12176 SV* properties = NULL; /* Code points that match \p{} \P{} */
12177 SV* posixes = NULL; /* Code points that match classes like, [:word:],
12178 extended beyond the Latin1 range */
12179 UV element_count = 0; /* Number of distinct elements in the class.
12180 Optimizations may be possible if this is tiny */
12181 AV * multi_char_matches = NULL; /* Code points that fold to more than one
12182 character; used under /i */
12184 char * stop_ptr = RExC_end; /* where to stop parsing */
12185 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
12187 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
12189 /* Unicode properties are stored in a swash; this holds the current one
12190 * being parsed. If this swash is the only above-latin1 component of the
12191 * character class, an optimization is to pass it directly on to the
12192 * execution engine. Otherwise, it is set to NULL to indicate that there
12193 * are other things in the class that have to be dealt with at execution
12195 SV* swash = NULL; /* Code points that match \p{} \P{} */
12197 /* Set if a component of this character class is user-defined; just passed
12198 * on to the engine */
12199 bool has_user_defined_property = FALSE;
12201 /* inversion list of code points this node matches only when the target
12202 * string is in UTF-8. (Because is under /d) */
12203 SV* depends_list = NULL;
12205 /* inversion list of code points this node matches. For much of the
12206 * function, it includes only those that match regardless of the utf8ness
12207 * of the target string */
12208 SV* cp_list = NULL;
12211 /* In a range, counts how many 0-2 of the ends of it came from literals,
12212 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
12213 UV literal_endpoint = 0;
12215 bool invert = FALSE; /* Is this class to be complemented */
12217 /* Is there any thing like \W or [:^digit:] that matches above the legal
12218 * Unicode range? */
12219 bool runtime_posix_matches_above_Unicode = FALSE;
12221 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
12222 case we need to change the emitted regop to an EXACT. */
12223 const char * orig_parse = RExC_parse;
12224 const I32 orig_size = RExC_size;
12225 GET_RE_DEBUG_FLAGS_DECL;
12227 PERL_ARGS_ASSERT_REGCLASS;
12229 PERL_UNUSED_ARG(depth);
12232 DEBUG_PARSE("clas");
12234 /* Assume we are going to generate an ANYOF node. */
12235 ret = reganode(pRExC_state, ANYOF, 0);
12238 RExC_size += ANYOF_SKIP;
12239 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
12242 ANYOF_FLAGS(ret) = 0;
12244 RExC_emit += ANYOF_SKIP;
12246 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
12248 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
12249 initial_listsv_len = SvCUR(listsv);
12250 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
12254 RExC_parse = regpatws(pRExC_state, RExC_parse,
12255 FALSE /* means don't recognize comments */);
12258 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
12261 allow_multi_folds = FALSE;
12264 RExC_parse = regpatws(pRExC_state, RExC_parse,
12265 FALSE /* means don't recognize comments */);
12269 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
12270 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
12271 const char *s = RExC_parse;
12272 const char c = *s++;
12274 while (isWORDCHAR(*s))
12276 if (*s && c == *s && s[1] == ']') {
12277 SAVEFREESV(RExC_rx_sv);
12279 "POSIX syntax [%c %c] belongs inside character classes",
12281 (void)ReREFCNT_inc(RExC_rx_sv);
12285 /* If the caller wants us to just parse a single element, accomplish this
12286 * by faking the loop ending condition */
12287 if (stop_at_1 && RExC_end > RExC_parse) {
12288 stop_ptr = RExC_parse + 1;
12291 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
12292 if (UCHARAT(RExC_parse) == ']')
12293 goto charclassloop;
12297 if (RExC_parse >= stop_ptr) {
12302 RExC_parse = regpatws(pRExC_state, RExC_parse,
12303 FALSE /* means don't recognize comments */);
12306 if (UCHARAT(RExC_parse) == ']') {
12312 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
12313 save_value = value;
12314 save_prevvalue = prevvalue;
12317 rangebegin = RExC_parse;
12321 value = utf8n_to_uvchr((U8*)RExC_parse,
12322 RExC_end - RExC_parse,
12323 &numlen, UTF8_ALLOW_DEFAULT);
12324 RExC_parse += numlen;
12327 value = UCHARAT(RExC_parse++);
12330 && RExC_parse < RExC_end
12331 && POSIXCC(UCHARAT(RExC_parse)))
12333 namedclass = regpposixcc(pRExC_state, value, strict);
12335 else if (value == '\\') {
12337 value = utf8n_to_uvchr((U8*)RExC_parse,
12338 RExC_end - RExC_parse,
12339 &numlen, UTF8_ALLOW_DEFAULT);
12340 RExC_parse += numlen;
12343 value = UCHARAT(RExC_parse++);
12345 /* Some compilers cannot handle switching on 64-bit integer
12346 * values, therefore value cannot be an UV. Yes, this will
12347 * be a problem later if we want switch on Unicode.
12348 * A similar issue a little bit later when switching on
12349 * namedclass. --jhi */
12351 /* If the \ is escaping white space when white space is being
12352 * skipped, it means that that white space is wanted literally, and
12353 * is already in 'value'. Otherwise, need to translate the escape
12354 * into what it signifies. */
12355 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
12357 case 'w': namedclass = ANYOF_WORDCHAR; break;
12358 case 'W': namedclass = ANYOF_NWORDCHAR; break;
12359 case 's': namedclass = ANYOF_SPACE; break;
12360 case 'S': namedclass = ANYOF_NSPACE; break;
12361 case 'd': namedclass = ANYOF_DIGIT; break;
12362 case 'D': namedclass = ANYOF_NDIGIT; break;
12363 case 'v': namedclass = ANYOF_VERTWS; break;
12364 case 'V': namedclass = ANYOF_NVERTWS; break;
12365 case 'h': namedclass = ANYOF_HORIZWS; break;
12366 case 'H': namedclass = ANYOF_NHORIZWS; break;
12367 case 'N': /* Handle \N{NAME} in class */
12369 /* We only pay attention to the first char of
12370 multichar strings being returned. I kinda wonder
12371 if this makes sense as it does change the behaviour
12372 from earlier versions, OTOH that behaviour was broken
12374 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
12375 TRUE, /* => charclass */
12378 if (*flagp & RESTART_UTF8)
12379 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12389 /* We will handle any undefined properties ourselves */
12390 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
12392 if (RExC_parse >= RExC_end)
12393 vFAIL2("Empty \\%c{}", (U8)value);
12394 if (*RExC_parse == '{') {
12395 const U8 c = (U8)value;
12396 e = strchr(RExC_parse++, '}');
12398 vFAIL2("Missing right brace on \\%c{}", c);
12399 while (isSPACE(UCHARAT(RExC_parse)))
12401 if (e == RExC_parse)
12402 vFAIL2("Empty \\%c{}", c);
12403 n = e - RExC_parse;
12404 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
12415 if (UCHARAT(RExC_parse) == '^') {
12418 /* toggle. (The rhs xor gets the single bit that
12419 * differs between P and p; the other xor inverts just
12421 value ^= 'P' ^ 'p';
12423 while (isSPACE(UCHARAT(RExC_parse))) {
12428 /* Try to get the definition of the property into
12429 * <invlist>. If /i is in effect, the effective property
12430 * will have its name be <__NAME_i>. The design is
12431 * discussed in commit
12432 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
12433 Newx(name, n + sizeof("_i__\n"), char);
12435 sprintf(name, "%s%.*s%s\n",
12436 (FOLD) ? "__" : "",
12442 /* Look up the property name, and get its swash and
12443 * inversion list, if the property is found */
12445 SvREFCNT_dec_NN(swash);
12447 swash = _core_swash_init("utf8", name, &PL_sv_undef,
12450 NULL, /* No inversion list */
12453 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
12455 SvREFCNT_dec_NN(swash);
12459 /* Here didn't find it. It could be a user-defined
12460 * property that will be available at run-time. If we
12461 * accept only compile-time properties, is an error;
12462 * otherwise add it to the list for run-time look up */
12464 RExC_parse = e + 1;
12465 vFAIL3("Property '%.*s' is unknown", (int) n, name);
12467 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
12468 (value == 'p' ? '+' : '!'),
12470 has_user_defined_property = TRUE;
12472 /* We don't know yet, so have to assume that the
12473 * property could match something in the Latin1 range,
12474 * hence something that isn't utf8. Note that this
12475 * would cause things in <depends_list> to match
12476 * inappropriately, except that any \p{}, including
12477 * this one forces Unicode semantics, which means there
12478 * is <no depends_list> */
12479 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
12483 /* Here, did get the swash and its inversion list. If
12484 * the swash is from a user-defined property, then this
12485 * whole character class should be regarded as such */
12486 has_user_defined_property =
12488 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
12490 /* Invert if asking for the complement */
12491 if (value == 'P') {
12492 _invlist_union_complement_2nd(properties,
12496 /* The swash can't be used as-is, because we've
12497 * inverted things; delay removing it to here after
12498 * have copied its invlist above */
12499 SvREFCNT_dec_NN(swash);
12503 _invlist_union(properties, invlist, &properties);
12508 RExC_parse = e + 1;
12509 namedclass = ANYOF_UNIPROP; /* no official name, but it's
12512 /* \p means they want Unicode semantics */
12513 RExC_uni_semantics = 1;
12516 case 'n': value = '\n'; break;
12517 case 'r': value = '\r'; break;
12518 case 't': value = '\t'; break;
12519 case 'f': value = '\f'; break;
12520 case 'b': value = '\b'; break;
12521 case 'e': value = ASCII_TO_NATIVE('\033');break;
12522 case 'a': value = ASCII_TO_NATIVE('\007');break;
12524 RExC_parse--; /* function expects to be pointed at the 'o' */
12526 const char* error_msg;
12527 bool valid = grok_bslash_o(&RExC_parse,
12530 SIZE_ONLY, /* warnings in pass
12533 silence_non_portable,
12539 if (PL_encoding && value < 0x100) {
12540 goto recode_encoding;
12544 RExC_parse--; /* function expects to be pointed at the 'x' */
12546 const char* error_msg;
12547 bool valid = grok_bslash_x(&RExC_parse,
12550 TRUE, /* Output warnings */
12552 silence_non_portable,
12558 if (PL_encoding && value < 0x100)
12559 goto recode_encoding;
12562 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
12564 case '0': case '1': case '2': case '3': case '4':
12565 case '5': case '6': case '7':
12567 /* Take 1-3 octal digits */
12568 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12569 numlen = (strict) ? 4 : 3;
12570 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
12571 RExC_parse += numlen;
12574 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12575 vFAIL("Need exactly 3 octal digits");
12577 else if (! SIZE_ONLY /* like \08, \178 */
12579 && RExC_parse < RExC_end
12580 && isDIGIT(*RExC_parse)
12581 && ckWARN(WARN_REGEXP))
12583 SAVEFREESV(RExC_rx_sv);
12584 reg_warn_non_literal_string(
12586 form_short_octal_warning(RExC_parse, numlen));
12587 (void)ReREFCNT_inc(RExC_rx_sv);
12590 if (PL_encoding && value < 0x100)
12591 goto recode_encoding;
12595 if (! RExC_override_recoding) {
12596 SV* enc = PL_encoding;
12597 value = reg_recode((const char)(U8)value, &enc);
12600 vFAIL("Invalid escape in the specified encoding");
12602 else if (SIZE_ONLY) {
12603 ckWARNreg(RExC_parse,
12604 "Invalid escape in the specified encoding");
12610 /* Allow \_ to not give an error */
12611 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
12613 vFAIL2("Unrecognized escape \\%c in character class",
12617 SAVEFREESV(RExC_rx_sv);
12618 ckWARN2reg(RExC_parse,
12619 "Unrecognized escape \\%c in character class passed through",
12621 (void)ReREFCNT_inc(RExC_rx_sv);
12625 } /* End of switch on char following backslash */
12626 } /* end of handling backslash escape sequences */
12629 literal_endpoint++;
12632 /* Here, we have the current token in 'value' */
12634 /* What matches in a locale is not known until runtime. This includes
12635 * what the Posix classes (like \w, [:space:]) match. Room must be
12636 * reserved (one time per class) to store such classes, either if Perl
12637 * is compiled so that locale nodes always should have this space, or
12638 * if there is such class info to be stored. The space will contain a
12639 * bit for each named class that is to be matched against. This isn't
12640 * needed for \p{} and pseudo-classes, as they are not affected by
12641 * locale, and hence are dealt with separately */
12644 && (ANYOF_LOCALE == ANYOF_CLASS
12645 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
12649 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12652 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12653 ANYOF_CLASS_ZERO(ret);
12655 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
12658 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
12660 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
12661 * literal, as is the character that began the false range, i.e.
12662 * the 'a' in the examples */
12665 const int w = (RExC_parse >= rangebegin)
12666 ? RExC_parse - rangebegin
12669 vFAIL4("False [] range \"%*.*s\"", w, w, rangebegin);
12672 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12673 ckWARN4reg(RExC_parse,
12674 "False [] range \"%*.*s\"",
12676 (void)ReREFCNT_inc(RExC_rx_sv);
12677 cp_list = add_cp_to_invlist(cp_list, '-');
12678 cp_list = add_cp_to_invlist(cp_list, prevvalue);
12682 range = 0; /* this was not a true range */
12683 element_count += 2; /* So counts for three values */
12687 U8 classnum = namedclass_to_classnum(namedclass);
12688 if (namedclass >= ANYOF_MAX) { /* If a special class */
12689 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
12691 /* Here, should be \h, \H, \v, or \V. Neither /d nor
12692 * /l make a difference in what these match. There
12693 * would be problems if these characters had folds
12694 * other than themselves, as cp_list is subject to
12696 if (classnum != _CC_VERTSPACE) {
12697 assert( namedclass == ANYOF_HORIZWS
12698 || namedclass == ANYOF_NHORIZWS);
12700 /* It turns out that \h is just a synonym for
12702 classnum = _CC_BLANK;
12705 _invlist_union_maybe_complement_2nd(
12707 PL_XPosix_ptrs[classnum],
12708 cBOOL(namedclass % 2), /* Complement if odd
12709 (NHORIZWS, NVERTWS)
12714 else if (classnum == _CC_ASCII) {
12717 ANYOF_CLASS_SET(ret, namedclass);
12720 #endif /* Not isascii(); just use the hard-coded definition for it */
12721 _invlist_union_maybe_complement_2nd(
12724 cBOOL(namedclass % 2), /* Complement if odd
12728 else { /* Garden variety class */
12730 /* The ascii range inversion list */
12731 SV* ascii_source = PL_Posix_ptrs[classnum];
12733 /* The full Latin1 range inversion list */
12734 SV* l1_source = PL_L1Posix_ptrs[classnum];
12736 /* This code is structured into two major clauses. The
12737 * first is for classes whose complete definitions may not
12738 * already be known. It not, the Latin1 definition
12739 * (guaranteed to already known) is used plus code is
12740 * generated to load the rest at run-time (only if needed).
12741 * If the complete definition is known, it drops down to
12742 * the second clause, where the complete definition is
12745 if (classnum < _FIRST_NON_SWASH_CC) {
12747 /* Here, the class has a swash, which may or not
12748 * already be loaded */
12750 /* The name of the property to use to match the full
12751 * eXtended Unicode range swash for this character
12753 const char *Xname = swash_property_names[classnum];
12755 /* If returning the inversion list, we can't defer
12756 * getting this until runtime */
12757 if (ret_invlist && ! PL_utf8_swash_ptrs[classnum]) {
12758 PL_utf8_swash_ptrs[classnum] =
12759 _core_swash_init("utf8", Xname, &PL_sv_undef,
12762 NULL, /* No inversion list */
12763 NULL /* No flags */
12765 assert(PL_utf8_swash_ptrs[classnum]);
12767 if ( ! PL_utf8_swash_ptrs[classnum]) {
12768 if (namedclass % 2 == 0) { /* A non-complemented
12770 /* If not /a matching, there are code points we
12771 * don't know at compile time. Arrange for the
12772 * unknown matches to be loaded at run-time, if
12774 if (! AT_LEAST_ASCII_RESTRICTED) {
12775 Perl_sv_catpvf(aTHX_ listsv, "+utf8::%s\n",
12778 if (LOC) { /* Under locale, set run-time
12780 ANYOF_CLASS_SET(ret, namedclass);
12783 /* Add the current class's code points to
12784 * the running total */
12785 _invlist_union(posixes,
12786 (AT_LEAST_ASCII_RESTRICTED)
12792 else { /* A complemented class */
12793 if (AT_LEAST_ASCII_RESTRICTED) {
12794 /* Under /a should match everything above
12795 * ASCII, plus the complement of the set's
12797 _invlist_union_complement_2nd(posixes,
12802 /* Arrange for the unknown matches to be
12803 * loaded at run-time, if needed */
12804 Perl_sv_catpvf(aTHX_ listsv, "!utf8::%s\n",
12806 runtime_posix_matches_above_Unicode = TRUE;
12808 ANYOF_CLASS_SET(ret, namedclass);
12812 /* We want to match everything in
12813 * Latin1, except those things that
12814 * l1_source matches */
12815 SV* scratch_list = NULL;
12816 _invlist_subtract(PL_Latin1, l1_source,
12819 /* Add the list from this class to the
12822 posixes = scratch_list;
12825 _invlist_union(posixes,
12828 SvREFCNT_dec_NN(scratch_list);
12830 if (DEPENDS_SEMANTICS) {
12832 |= ANYOF_NON_UTF8_LATIN1_ALL;
12837 goto namedclass_done;
12840 /* Here, there is a swash loaded for the class. If no
12841 * inversion list for it yet, get it */
12842 if (! PL_XPosix_ptrs[classnum]) {
12843 PL_XPosix_ptrs[classnum]
12844 = _swash_to_invlist(PL_utf8_swash_ptrs[classnum]);
12848 /* Here there is an inversion list already loaded for the
12851 if (namedclass % 2 == 0) { /* A non-complemented class,
12852 like ANYOF_PUNCT */
12854 /* For non-locale, just add it to any existing list
12856 _invlist_union(posixes,
12857 (AT_LEAST_ASCII_RESTRICTED)
12859 : PL_XPosix_ptrs[classnum],
12862 else { /* Locale */
12863 SV* scratch_list = NULL;
12865 /* For above Latin1 code points, we use the full
12867 _invlist_intersection(PL_AboveLatin1,
12868 PL_XPosix_ptrs[classnum],
12870 /* And set the output to it, adding instead if
12871 * there already is an output. Checking if
12872 * 'posixes' is NULL first saves an extra clone.
12873 * Its reference count will be decremented at the
12874 * next union, etc, or if this is the only
12875 * instance, at the end of the routine */
12877 posixes = scratch_list;
12880 _invlist_union(posixes, scratch_list, &posixes);
12881 SvREFCNT_dec_NN(scratch_list);
12884 #ifndef HAS_ISBLANK
12885 if (namedclass != ANYOF_BLANK) {
12887 /* Set this class in the node for runtime
12889 ANYOF_CLASS_SET(ret, namedclass);
12890 #ifndef HAS_ISBLANK
12893 /* No isblank(), use the hard-coded ASCII-range
12894 * blanks, adding them to the running total. */
12896 _invlist_union(posixes, ascii_source, &posixes);
12901 else { /* A complemented class, like ANYOF_NPUNCT */
12903 _invlist_union_complement_2nd(
12905 (AT_LEAST_ASCII_RESTRICTED)
12907 : PL_XPosix_ptrs[classnum],
12909 /* Under /d, everything in the upper half of the
12910 * Latin1 range matches this complement */
12911 if (DEPENDS_SEMANTICS) {
12912 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12915 else { /* Locale */
12916 SV* scratch_list = NULL;
12917 _invlist_subtract(PL_AboveLatin1,
12918 PL_XPosix_ptrs[classnum],
12921 posixes = scratch_list;
12924 _invlist_union(posixes, scratch_list, &posixes);
12925 SvREFCNT_dec_NN(scratch_list);
12927 #ifndef HAS_ISBLANK
12928 if (namedclass != ANYOF_NBLANK) {
12930 ANYOF_CLASS_SET(ret, namedclass);
12931 #ifndef HAS_ISBLANK
12934 /* Get the list of all code points in Latin1
12935 * that are not ASCII blanks, and add them to
12936 * the running total */
12937 _invlist_subtract(PL_Latin1, ascii_source,
12939 _invlist_union(posixes, scratch_list, &posixes);
12940 SvREFCNT_dec_NN(scratch_list);
12947 continue; /* Go get next character */
12949 } /* end of namedclass \blah */
12951 /* Here, we have a single value. If 'range' is set, it is the ending
12952 * of a range--check its validity. Later, we will handle each
12953 * individual code point in the range. If 'range' isn't set, this
12954 * could be the beginning of a range, so check for that by looking
12955 * ahead to see if the next real character to be processed is the range
12956 * indicator--the minus sign */
12959 RExC_parse = regpatws(pRExC_state, RExC_parse,
12960 FALSE /* means don't recognize comments */);
12964 if (prevvalue > value) /* b-a */ {
12965 const int w = RExC_parse - rangebegin;
12966 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12967 range = 0; /* not a valid range */
12971 prevvalue = value; /* save the beginning of the potential range */
12972 if (! stop_at_1 /* Can't be a range if parsing just one thing */
12973 && *RExC_parse == '-')
12975 char* next_char_ptr = RExC_parse + 1;
12976 if (skip_white) { /* Get the next real char after the '-' */
12977 next_char_ptr = regpatws(pRExC_state,
12979 FALSE); /* means don't recognize
12983 /* If the '-' is at the end of the class (just before the ']',
12984 * it is a literal minus; otherwise it is a range */
12985 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
12986 RExC_parse = next_char_ptr;
12988 /* a bad range like \w-, [:word:]- ? */
12989 if (namedclass > OOB_NAMEDCLASS) {
12990 if (strict || ckWARN(WARN_REGEXP)) {
12992 RExC_parse >= rangebegin ?
12993 RExC_parse - rangebegin : 0;
12995 vFAIL4("False [] range \"%*.*s\"",
13000 "False [] range \"%*.*s\"",
13005 cp_list = add_cp_to_invlist(cp_list, '-');
13009 range = 1; /* yeah, it's a range! */
13010 continue; /* but do it the next time */
13015 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
13018 /* non-Latin1 code point implies unicode semantics. Must be set in
13019 * pass1 so is there for the whole of pass 2 */
13021 RExC_uni_semantics = 1;
13024 /* Ready to process either the single value, or the completed range.
13025 * For single-valued non-inverted ranges, we consider the possibility
13026 * of multi-char folds. (We made a conscious decision to not do this
13027 * for the other cases because it can often lead to non-intuitive
13028 * results. For example, you have the peculiar case that:
13029 * "s s" =~ /^[^\xDF]+$/i => Y
13030 * "ss" =~ /^[^\xDF]+$/i => N
13032 * See [perl #89750] */
13033 if (FOLD && allow_multi_folds && value == prevvalue) {
13034 if (value == LATIN_SMALL_LETTER_SHARP_S
13035 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
13038 /* Here <value> is indeed a multi-char fold. Get what it is */
13040 U8 foldbuf[UTF8_MAXBYTES_CASE];
13043 UV folded = _to_uni_fold_flags(
13048 | ((LOC) ? FOLD_FLAGS_LOCALE
13049 : (ASCII_FOLD_RESTRICTED)
13050 ? FOLD_FLAGS_NOMIX_ASCII
13054 /* Here, <folded> should be the first character of the
13055 * multi-char fold of <value>, with <foldbuf> containing the
13056 * whole thing. But, if this fold is not allowed (because of
13057 * the flags), <fold> will be the same as <value>, and should
13058 * be processed like any other character, so skip the special
13060 if (folded != value) {
13062 /* Skip if we are recursed, currently parsing the class
13063 * again. Otherwise add this character to the list of
13064 * multi-char folds. */
13065 if (! RExC_in_multi_char_class) {
13066 AV** this_array_ptr;
13068 STRLEN cp_count = utf8_length(foldbuf,
13069 foldbuf + foldlen);
13070 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
13072 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
13075 if (! multi_char_matches) {
13076 multi_char_matches = newAV();
13079 /* <multi_char_matches> is actually an array of arrays.
13080 * There will be one or two top-level elements: [2],
13081 * and/or [3]. The [2] element is an array, each
13082 * element thereof is a character which folds to TWO
13083 * characters; [3] is for folds to THREE characters.
13084 * (Unicode guarantees a maximum of 3 characters in any
13085 * fold.) When we rewrite the character class below,
13086 * we will do so such that the longest folds are
13087 * written first, so that it prefers the longest
13088 * matching strings first. This is done even if it
13089 * turns out that any quantifier is non-greedy, out of
13090 * programmer laziness. Tom Christiansen has agreed
13091 * that this is ok. This makes the test for the
13092 * ligature 'ffi' come before the test for 'ff' */
13093 if (av_exists(multi_char_matches, cp_count)) {
13094 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13096 this_array = *this_array_ptr;
13099 this_array = newAV();
13100 av_store(multi_char_matches, cp_count,
13103 av_push(this_array, multi_fold);
13106 /* This element should not be processed further in this
13109 value = save_value;
13110 prevvalue = save_prevvalue;
13116 /* Deal with this element of the class */
13119 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
13121 SV* this_range = _new_invlist(1);
13122 _append_range_to_invlist(this_range, prevvalue, value);
13124 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
13125 * If this range was specified using something like 'i-j', we want
13126 * to include only the 'i' and the 'j', and not anything in
13127 * between, so exclude non-ASCII, non-alphabetics from it.
13128 * However, if the range was specified with something like
13129 * [\x89-\x91] or [\x89-j], all code points within it should be
13130 * included. literal_endpoint==2 means both ends of the range used
13131 * a literal character, not \x{foo} */
13132 if (literal_endpoint == 2
13133 && (prevvalue >= 'a' && value <= 'z')
13134 || (prevvalue >= 'A' && value <= 'Z'))
13136 _invlist_intersection(this_range, PL_Posix_ptrs[_CC_ALPHA],
13139 _invlist_union(cp_list, this_range, &cp_list);
13140 literal_endpoint = 0;
13144 range = 0; /* this range (if it was one) is done now */
13145 } /* End of loop through all the text within the brackets */
13147 /* If anything in the class expands to more than one character, we have to
13148 * deal with them by building up a substitute parse string, and recursively
13149 * calling reg() on it, instead of proceeding */
13150 if (multi_char_matches) {
13151 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
13154 char *save_end = RExC_end;
13155 char *save_parse = RExC_parse;
13156 bool first_time = TRUE; /* First multi-char occurrence doesn't get
13161 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
13162 because too confusing */
13164 sv_catpv(substitute_parse, "(?:");
13168 /* Look at the longest folds first */
13169 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
13171 if (av_exists(multi_char_matches, cp_count)) {
13172 AV** this_array_ptr;
13175 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13177 while ((this_sequence = av_pop(*this_array_ptr)) !=
13180 if (! first_time) {
13181 sv_catpv(substitute_parse, "|");
13183 first_time = FALSE;
13185 sv_catpv(substitute_parse, SvPVX(this_sequence));
13190 /* If the character class contains anything else besides these
13191 * multi-character folds, have to include it in recursive parsing */
13192 if (element_count) {
13193 sv_catpv(substitute_parse, "|[");
13194 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
13195 sv_catpv(substitute_parse, "]");
13198 sv_catpv(substitute_parse, ")");
13201 /* This is a way to get the parse to skip forward a whole named
13202 * sequence instead of matching the 2nd character when it fails the
13204 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
13208 RExC_parse = SvPV(substitute_parse, len);
13209 RExC_end = RExC_parse + len;
13210 RExC_in_multi_char_class = 1;
13211 RExC_emit = (regnode *)orig_emit;
13213 ret = reg(pRExC_state, 1, ®_flags, depth+1);
13215 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
13217 RExC_parse = save_parse;
13218 RExC_end = save_end;
13219 RExC_in_multi_char_class = 0;
13220 SvREFCNT_dec_NN(multi_char_matches);
13224 /* If the character class contains only a single element, it may be
13225 * optimizable into another node type which is smaller and runs faster.
13226 * Check if this is the case for this class */
13227 if (element_count == 1 && ! ret_invlist) {
13231 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
13232 [:digit:] or \p{foo} */
13234 /* All named classes are mapped into POSIXish nodes, with its FLAG
13235 * argument giving which class it is */
13236 switch ((I32)namedclass) {
13237 case ANYOF_UNIPROP:
13240 /* These don't depend on the charset modifiers. They always
13241 * match under /u rules */
13242 case ANYOF_NHORIZWS:
13243 case ANYOF_HORIZWS:
13244 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
13247 case ANYOF_NVERTWS:
13252 /* The actual POSIXish node for all the rest depends on the
13253 * charset modifier. The ones in the first set depend only on
13254 * ASCII or, if available on this platform, locale */
13258 op = (LOC) ? POSIXL : POSIXA;
13269 /* under /a could be alpha */
13271 if (ASCII_RESTRICTED) {
13272 namedclass = ANYOF_ALPHA + (namedclass % 2);
13280 /* The rest have more possibilities depending on the charset.
13281 * We take advantage of the enum ordering of the charset
13282 * modifiers to get the exact node type, */
13284 op = POSIXD + get_regex_charset(RExC_flags);
13285 if (op > POSIXA) { /* /aa is same as /a */
13288 #ifndef HAS_ISBLANK
13290 && (namedclass == ANYOF_BLANK
13291 || namedclass == ANYOF_NBLANK))
13298 /* The odd numbered ones are the complements of the
13299 * next-lower even number one */
13300 if (namedclass % 2 == 1) {
13304 arg = namedclass_to_classnum(namedclass);
13308 else if (value == prevvalue) {
13310 /* Here, the class consists of just a single code point */
13313 if (! LOC && value == '\n') {
13314 op = REG_ANY; /* Optimize [^\n] */
13315 *flagp |= HASWIDTH|SIMPLE;
13319 else if (value < 256 || UTF) {
13321 /* Optimize a single value into an EXACTish node, but not if it
13322 * would require converting the pattern to UTF-8. */
13323 op = compute_EXACTish(pRExC_state);
13325 } /* Otherwise is a range */
13326 else if (! LOC) { /* locale could vary these */
13327 if (prevvalue == '0') {
13328 if (value == '9') {
13335 /* Here, we have changed <op> away from its initial value iff we found
13336 * an optimization */
13339 /* Throw away this ANYOF regnode, and emit the calculated one,
13340 * which should correspond to the beginning, not current, state of
13342 const char * cur_parse = RExC_parse;
13343 RExC_parse = (char *)orig_parse;
13347 /* To get locale nodes to not use the full ANYOF size would
13348 * require moving the code above that writes the portions
13349 * of it that aren't in other nodes to after this point.
13350 * e.g. ANYOF_CLASS_SET */
13351 RExC_size = orig_size;
13355 RExC_emit = (regnode *)orig_emit;
13356 if (PL_regkind[op] == POSIXD) {
13358 op += NPOSIXD - POSIXD;
13363 ret = reg_node(pRExC_state, op);
13365 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
13369 *flagp |= HASWIDTH|SIMPLE;
13371 else if (PL_regkind[op] == EXACT) {
13372 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13375 RExC_parse = (char *) cur_parse;
13377 SvREFCNT_dec(posixes);
13378 SvREFCNT_dec(cp_list);
13385 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
13387 /* If folding, we calculate all characters that could fold to or from the
13388 * ones already on the list */
13389 if (FOLD && cp_list) {
13390 UV start, end; /* End points of code point ranges */
13392 SV* fold_intersection = NULL;
13394 /* If the highest code point is within Latin1, we can use the
13395 * compiled-in Alphas list, and not have to go out to disk. This
13396 * yields two false positives, the masculine and feminine ordinal
13397 * indicators, which are weeded out below using the
13398 * IS_IN_SOME_FOLD_L1() macro */
13399 if (invlist_highest(cp_list) < 256) {
13400 _invlist_intersection(PL_L1Posix_ptrs[_CC_ALPHA], cp_list,
13401 &fold_intersection);
13405 /* Here, there are non-Latin1 code points, so we will have to go
13406 * fetch the list of all the characters that participate in folds
13408 if (! PL_utf8_foldable) {
13409 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13410 &PL_sv_undef, 1, 0);
13411 PL_utf8_foldable = _get_swash_invlist(swash);
13412 SvREFCNT_dec_NN(swash);
13415 /* This is a hash that for a particular fold gives all characters
13416 * that are involved in it */
13417 if (! PL_utf8_foldclosures) {
13419 /* If we were unable to find any folds, then we likely won't be
13420 * able to find the closures. So just create an empty list.
13421 * Folding will effectively be restricted to the non-Unicode
13422 * rules hard-coded into Perl. (This case happens legitimately
13423 * during compilation of Perl itself before the Unicode tables
13424 * are generated) */
13425 if (_invlist_len(PL_utf8_foldable) == 0) {
13426 PL_utf8_foldclosures = newHV();
13429 /* If the folds haven't been read in, call a fold function
13431 if (! PL_utf8_tofold) {
13432 U8 dummy[UTF8_MAXBYTES+1];
13434 /* This string is just a short named one above \xff */
13435 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
13436 assert(PL_utf8_tofold); /* Verify that worked */
13438 PL_utf8_foldclosures =
13439 _swash_inversion_hash(PL_utf8_tofold);
13443 /* Only the characters in this class that participate in folds need
13444 * be checked. Get the intersection of this class and all the
13445 * possible characters that are foldable. This can quickly narrow
13446 * down a large class */
13447 _invlist_intersection(PL_utf8_foldable, cp_list,
13448 &fold_intersection);
13451 /* Now look at the foldable characters in this class individually */
13452 invlist_iterinit(fold_intersection);
13453 while (invlist_iternext(fold_intersection, &start, &end)) {
13456 /* Locale folding for Latin1 characters is deferred until runtime */
13457 if (LOC && start < 256) {
13461 /* Look at every character in the range */
13462 for (j = start; j <= end; j++) {
13464 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
13470 /* We have the latin1 folding rules hard-coded here so that
13471 * an innocent-looking character class, like /[ks]/i won't
13472 * have to go out to disk to find the possible matches.
13473 * XXX It would be better to generate these via regen, in
13474 * case a new version of the Unicode standard adds new
13475 * mappings, though that is not really likely, and may be
13476 * caught by the default: case of the switch below. */
13478 if (IS_IN_SOME_FOLD_L1(j)) {
13480 /* ASCII is always matched; non-ASCII is matched only
13481 * under Unicode rules */
13482 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
13484 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
13488 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
13492 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
13493 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
13495 /* Certain Latin1 characters have matches outside
13496 * Latin1. To get here, <j> is one of those
13497 * characters. None of these matches is valid for
13498 * ASCII characters under /aa, which is why the 'if'
13499 * just above excludes those. These matches only
13500 * happen when the target string is utf8. The code
13501 * below adds the single fold closures for <j> to the
13502 * inversion list. */
13507 add_cp_to_invlist(cp_list, KELVIN_SIGN);
13511 cp_list = add_cp_to_invlist(cp_list,
13512 LATIN_SMALL_LETTER_LONG_S);
13515 cp_list = add_cp_to_invlist(cp_list,
13516 GREEK_CAPITAL_LETTER_MU);
13517 cp_list = add_cp_to_invlist(cp_list,
13518 GREEK_SMALL_LETTER_MU);
13520 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13521 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13523 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
13525 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13526 cp_list = add_cp_to_invlist(cp_list,
13527 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13529 case LATIN_SMALL_LETTER_SHARP_S:
13530 cp_list = add_cp_to_invlist(cp_list,
13531 LATIN_CAPITAL_LETTER_SHARP_S);
13533 case 'F': case 'f':
13534 case 'I': case 'i':
13535 case 'L': case 'l':
13536 case 'T': case 't':
13537 case 'A': case 'a':
13538 case 'H': case 'h':
13539 case 'J': case 'j':
13540 case 'N': case 'n':
13541 case 'W': case 'w':
13542 case 'Y': case 'y':
13543 /* These all are targets of multi-character
13544 * folds from code points that require UTF8 to
13545 * express, so they can't match unless the
13546 * target string is in UTF-8, so no action here
13547 * is necessary, as regexec.c properly handles
13548 * the general case for UTF-8 matching and
13549 * multi-char folds */
13552 /* Use deprecated warning to increase the
13553 * chances of this being output */
13554 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
13561 /* Here is an above Latin1 character. We don't have the rules
13562 * hard-coded for it. First, get its fold. This is the simple
13563 * fold, as the multi-character folds have been handled earlier
13564 * and separated out */
13565 _to_uni_fold_flags(j, foldbuf, &foldlen,
13567 ? FOLD_FLAGS_LOCALE
13568 : (ASCII_FOLD_RESTRICTED)
13569 ? FOLD_FLAGS_NOMIX_ASCII
13572 /* Single character fold of above Latin1. Add everything in
13573 * its fold closure to the list that this node should match.
13574 * The fold closures data structure is a hash with the keys
13575 * being the UTF-8 of every character that is folded to, like
13576 * 'k', and the values each an array of all code points that
13577 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
13578 * Multi-character folds are not included */
13579 if ((listp = hv_fetch(PL_utf8_foldclosures,
13580 (char *) foldbuf, foldlen, FALSE)))
13582 AV* list = (AV*) *listp;
13584 for (k = 0; k <= av_len(list); k++) {
13585 SV** c_p = av_fetch(list, k, FALSE);
13588 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
13592 /* /aa doesn't allow folds between ASCII and non-; /l
13593 * doesn't allow them between above and below 256 */
13594 if ((ASCII_FOLD_RESTRICTED
13595 && (isASCII(c) != isASCII(j)))
13596 || (LOC && c < 256)) {
13600 /* Folds involving non-ascii Latin1 characters
13601 * under /d are added to a separate list */
13602 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
13604 cp_list = add_cp_to_invlist(cp_list, c);
13607 depends_list = add_cp_to_invlist(depends_list, c);
13613 SvREFCNT_dec_NN(fold_intersection);
13616 /* And combine the result (if any) with any inversion list from posix
13617 * classes. The lists are kept separate up to now because we don't want to
13618 * fold the classes (folding of those is automatically handled by the swash
13619 * fetching code) */
13621 if (! DEPENDS_SEMANTICS) {
13623 _invlist_union(cp_list, posixes, &cp_list);
13624 SvREFCNT_dec_NN(posixes);
13631 /* Under /d, we put into a separate list the Latin1 things that
13632 * match only when the target string is utf8 */
13633 SV* nonascii_but_latin1_properties = NULL;
13634 _invlist_intersection(posixes, PL_Latin1,
13635 &nonascii_but_latin1_properties);
13636 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
13637 &nonascii_but_latin1_properties);
13638 _invlist_subtract(posixes, nonascii_but_latin1_properties,
13641 _invlist_union(cp_list, posixes, &cp_list);
13642 SvREFCNT_dec_NN(posixes);
13648 if (depends_list) {
13649 _invlist_union(depends_list, nonascii_but_latin1_properties,
13651 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
13654 depends_list = nonascii_but_latin1_properties;
13659 /* And combine the result (if any) with any inversion list from properties.
13660 * The lists are kept separate up to now so that we can distinguish the two
13661 * in regards to matching above-Unicode. A run-time warning is generated
13662 * if a Unicode property is matched against a non-Unicode code point. But,
13663 * we allow user-defined properties to match anything, without any warning,
13664 * and we also suppress the warning if there is a portion of the character
13665 * class that isn't a Unicode property, and which matches above Unicode, \W
13666 * or [\x{110000}] for example.
13667 * (Note that in this case, unlike the Posix one above, there is no
13668 * <depends_list>, because having a Unicode property forces Unicode
13671 bool warn_super = ! has_user_defined_property;
13674 /* If it matters to the final outcome, see if a non-property
13675 * component of the class matches above Unicode. If so, the
13676 * warning gets suppressed. This is true even if just a single
13677 * such code point is specified, as though not strictly correct if
13678 * another such code point is matched against, the fact that they
13679 * are using above-Unicode code points indicates they should know
13680 * the issues involved */
13682 bool non_prop_matches_above_Unicode =
13683 runtime_posix_matches_above_Unicode
13684 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
13686 non_prop_matches_above_Unicode =
13687 ! non_prop_matches_above_Unicode;
13689 warn_super = ! non_prop_matches_above_Unicode;
13692 _invlist_union(properties, cp_list, &cp_list);
13693 SvREFCNT_dec_NN(properties);
13696 cp_list = properties;
13700 OP(ret) = ANYOF_WARN_SUPER;
13704 /* Here, we have calculated what code points should be in the character
13707 * Now we can see about various optimizations. Fold calculation (which we
13708 * did above) needs to take place before inversion. Otherwise /[^k]/i
13709 * would invert to include K, which under /i would match k, which it
13710 * shouldn't. Therefore we can't invert folded locale now, as it won't be
13711 * folded until runtime */
13713 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
13714 * at compile time. Besides not inverting folded locale now, we can't
13715 * invert if there are things such as \w, which aren't known until runtime
13718 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
13720 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13722 _invlist_invert(cp_list);
13724 /* Any swash can't be used as-is, because we've inverted things */
13726 SvREFCNT_dec_NN(swash);
13730 /* Clear the invert flag since have just done it here */
13735 *ret_invlist = cp_list;
13736 SvREFCNT_dec(swash);
13738 /* Discard the generated node */
13740 RExC_size = orig_size;
13743 RExC_emit = orig_emit;
13748 /* If we didn't do folding, it's because some information isn't available
13749 * until runtime; set the run-time fold flag for these. (We don't have to
13750 * worry about properties folding, as that is taken care of by the swash
13754 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
13757 /* Some character classes are equivalent to other nodes. Such nodes take
13758 * up less room and generally fewer operations to execute than ANYOF nodes.
13759 * Above, we checked for and optimized into some such equivalents for
13760 * certain common classes that are easy to test. Getting to this point in
13761 * the code means that the class didn't get optimized there. Since this
13762 * code is only executed in Pass 2, it is too late to save space--it has
13763 * been allocated in Pass 1, and currently isn't given back. But turning
13764 * things into an EXACTish node can allow the optimizer to join it to any
13765 * adjacent such nodes. And if the class is equivalent to things like /./,
13766 * expensive run-time swashes can be avoided. Now that we have more
13767 * complete information, we can find things necessarily missed by the
13768 * earlier code. I (khw) am not sure how much to look for here. It would
13769 * be easy, but perhaps too slow, to check any candidates against all the
13770 * node types they could possibly match using _invlistEQ(). */
13775 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
13776 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13779 U8 op = END; /* The optimzation node-type */
13780 const char * cur_parse= RExC_parse;
13782 invlist_iterinit(cp_list);
13783 if (! invlist_iternext(cp_list, &start, &end)) {
13785 /* Here, the list is empty. This happens, for example, when a
13786 * Unicode property is the only thing in the character class, and
13787 * it doesn't match anything. (perluniprops.pod notes such
13790 *flagp |= HASWIDTH|SIMPLE;
13792 else if (start == end) { /* The range is a single code point */
13793 if (! invlist_iternext(cp_list, &start, &end)
13795 /* Don't do this optimization if it would require changing
13796 * the pattern to UTF-8 */
13797 && (start < 256 || UTF))
13799 /* Here, the list contains a single code point. Can optimize
13800 * into an EXACT node */
13809 /* A locale node under folding with one code point can be
13810 * an EXACTFL, as its fold won't be calculated until
13816 /* Here, we are generally folding, but there is only one
13817 * code point to match. If we have to, we use an EXACT
13818 * node, but it would be better for joining with adjacent
13819 * nodes in the optimization pass if we used the same
13820 * EXACTFish node that any such are likely to be. We can
13821 * do this iff the code point doesn't participate in any
13822 * folds. For example, an EXACTF of a colon is the same as
13823 * an EXACT one, since nothing folds to or from a colon. */
13825 if (IS_IN_SOME_FOLD_L1(value)) {
13830 if (! PL_utf8_foldable) {
13831 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13832 &PL_sv_undef, 1, 0);
13833 PL_utf8_foldable = _get_swash_invlist(swash);
13834 SvREFCNT_dec_NN(swash);
13836 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
13841 /* If we haven't found the node type, above, it means we
13842 * can use the prevailing one */
13844 op = compute_EXACTish(pRExC_state);
13849 else if (start == 0) {
13850 if (end == UV_MAX) {
13852 *flagp |= HASWIDTH|SIMPLE;
13855 else if (end == '\n' - 1
13856 && invlist_iternext(cp_list, &start, &end)
13857 && start == '\n' + 1 && end == UV_MAX)
13860 *flagp |= HASWIDTH|SIMPLE;
13864 invlist_iterfinish(cp_list);
13867 RExC_parse = (char *)orig_parse;
13868 RExC_emit = (regnode *)orig_emit;
13870 ret = reg_node(pRExC_state, op);
13872 RExC_parse = (char *)cur_parse;
13874 if (PL_regkind[op] == EXACT) {
13875 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13878 SvREFCNT_dec_NN(cp_list);
13883 /* Here, <cp_list> contains all the code points we can determine at
13884 * compile time that match under all conditions. Go through it, and
13885 * for things that belong in the bitmap, put them there, and delete from
13886 * <cp_list>. While we are at it, see if everything above 255 is in the
13887 * list, and if so, set a flag to speed up execution */
13888 ANYOF_BITMAP_ZERO(ret);
13891 /* This gets set if we actually need to modify things */
13892 bool change_invlist = FALSE;
13896 /* Start looking through <cp_list> */
13897 invlist_iterinit(cp_list);
13898 while (invlist_iternext(cp_list, &start, &end)) {
13902 if (end == UV_MAX && start <= 256) {
13903 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
13906 /* Quit if are above what we should change */
13911 change_invlist = TRUE;
13913 /* Set all the bits in the range, up to the max that we are doing */
13914 high = (end < 255) ? end : 255;
13915 for (i = start; i <= (int) high; i++) {
13916 if (! ANYOF_BITMAP_TEST(ret, i)) {
13917 ANYOF_BITMAP_SET(ret, i);
13921 invlist_iterfinish(cp_list);
13923 /* Done with loop; remove any code points that are in the bitmap from
13925 if (change_invlist) {
13926 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
13929 /* If have completely emptied it, remove it completely */
13930 if (_invlist_len(cp_list) == 0) {
13931 SvREFCNT_dec_NN(cp_list);
13937 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
13940 /* Here, the bitmap has been populated with all the Latin1 code points that
13941 * always match. Can now add to the overall list those that match only
13942 * when the target string is UTF-8 (<depends_list>). */
13943 if (depends_list) {
13945 _invlist_union(cp_list, depends_list, &cp_list);
13946 SvREFCNT_dec_NN(depends_list);
13949 cp_list = depends_list;
13953 /* If there is a swash and more than one element, we can't use the swash in
13954 * the optimization below. */
13955 if (swash && element_count > 1) {
13956 SvREFCNT_dec_NN(swash);
13961 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13963 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
13966 /* av[0] stores the character class description in its textual form:
13967 * used later (regexec.c:Perl_regclass_swash()) to initialize the
13968 * appropriate swash, and is also useful for dumping the regnode.
13969 * av[1] if NULL, is a placeholder to later contain the swash computed
13970 * from av[0]. But if no further computation need be done, the
13971 * swash is stored there now.
13972 * av[2] stores the cp_list inversion list for use in addition or
13973 * instead of av[0]; used only if av[1] is NULL
13974 * av[3] is set if any component of the class is from a user-defined
13975 * property; used only if av[1] is NULL */
13976 AV * const av = newAV();
13979 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13980 ? SvREFCNT_inc(listsv) : &PL_sv_undef);
13982 av_store(av, 1, swash);
13983 SvREFCNT_dec_NN(cp_list);
13986 av_store(av, 1, NULL);
13988 av_store(av, 2, cp_list);
13989 av_store(av, 3, newSVuv(has_user_defined_property));
13993 rv = newRV_noinc(MUTABLE_SV(av));
13994 n = add_data(pRExC_state, 1, "s");
13995 RExC_rxi->data->data[n] = (void*)rv;
13999 *flagp |= HASWIDTH|SIMPLE;
14002 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14005 /* reg_skipcomment()
14007 Absorbs an /x style # comments from the input stream.
14008 Returns true if there is more text remaining in the stream.
14009 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
14010 terminates the pattern without including a newline.
14012 Note its the callers responsibility to ensure that we are
14013 actually in /x mode
14018 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
14022 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
14024 while (RExC_parse < RExC_end)
14025 if (*RExC_parse++ == '\n') {
14030 /* we ran off the end of the pattern without ending
14031 the comment, so we have to add an \n when wrapping */
14032 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
14040 Advances the parse position, and optionally absorbs
14041 "whitespace" from the inputstream.
14043 Without /x "whitespace" means (?#...) style comments only,
14044 with /x this means (?#...) and # comments and whitespace proper.
14046 Returns the RExC_parse point from BEFORE the scan occurs.
14048 This is the /x friendly way of saying RExC_parse++.
14052 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
14054 char* const retval = RExC_parse++;
14056 PERL_ARGS_ASSERT_NEXTCHAR;
14059 if (RExC_end - RExC_parse >= 3
14060 && *RExC_parse == '('
14061 && RExC_parse[1] == '?'
14062 && RExC_parse[2] == '#')
14064 while (*RExC_parse != ')') {
14065 if (RExC_parse == RExC_end)
14066 FAIL("Sequence (?#... not terminated");
14072 if (RExC_flags & RXf_PMf_EXTENDED) {
14073 if (isSPACE(*RExC_parse)) {
14077 else if (*RExC_parse == '#') {
14078 if ( reg_skipcomment( pRExC_state ) )
14087 - reg_node - emit a node
14089 STATIC regnode * /* Location. */
14090 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
14094 regnode * const ret = RExC_emit;
14095 GET_RE_DEBUG_FLAGS_DECL;
14097 PERL_ARGS_ASSERT_REG_NODE;
14100 SIZE_ALIGN(RExC_size);
14104 if (RExC_emit >= RExC_emit_bound)
14105 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14106 op, RExC_emit, RExC_emit_bound);
14108 NODE_ALIGN_FILL(ret);
14110 FILL_ADVANCE_NODE(ptr, op);
14111 #ifdef RE_TRACK_PATTERN_OFFSETS
14112 if (RExC_offsets) { /* MJD */
14113 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
14114 "reg_node", __LINE__,
14116 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
14117 ? "Overwriting end of array!\n" : "OK",
14118 (UV)(RExC_emit - RExC_emit_start),
14119 (UV)(RExC_parse - RExC_start),
14120 (UV)RExC_offsets[0]));
14121 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
14129 - reganode - emit a node with an argument
14131 STATIC regnode * /* Location. */
14132 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
14136 regnode * const ret = RExC_emit;
14137 GET_RE_DEBUG_FLAGS_DECL;
14139 PERL_ARGS_ASSERT_REGANODE;
14142 SIZE_ALIGN(RExC_size);
14147 assert(2==regarglen[op]+1);
14149 Anything larger than this has to allocate the extra amount.
14150 If we changed this to be:
14152 RExC_size += (1 + regarglen[op]);
14154 then it wouldn't matter. Its not clear what side effect
14155 might come from that so its not done so far.
14160 if (RExC_emit >= RExC_emit_bound)
14161 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14162 op, RExC_emit, RExC_emit_bound);
14164 NODE_ALIGN_FILL(ret);
14166 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
14167 #ifdef RE_TRACK_PATTERN_OFFSETS
14168 if (RExC_offsets) { /* MJD */
14169 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14173 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
14174 "Overwriting end of array!\n" : "OK",
14175 (UV)(RExC_emit - RExC_emit_start),
14176 (UV)(RExC_parse - RExC_start),
14177 (UV)RExC_offsets[0]));
14178 Set_Cur_Node_Offset;
14186 - reguni - emit (if appropriate) a Unicode character
14189 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
14193 PERL_ARGS_ASSERT_REGUNI;
14195 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
14199 - reginsert - insert an operator in front of already-emitted operand
14201 * Means relocating the operand.
14204 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
14210 const int offset = regarglen[(U8)op];
14211 const int size = NODE_STEP_REGNODE + offset;
14212 GET_RE_DEBUG_FLAGS_DECL;
14214 PERL_ARGS_ASSERT_REGINSERT;
14215 PERL_UNUSED_ARG(depth);
14216 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
14217 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
14226 if (RExC_open_parens) {
14228 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
14229 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
14230 if ( RExC_open_parens[paren] >= opnd ) {
14231 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
14232 RExC_open_parens[paren] += size;
14234 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
14236 if ( RExC_close_parens[paren] >= opnd ) {
14237 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
14238 RExC_close_parens[paren] += size;
14240 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
14245 while (src > opnd) {
14246 StructCopy(--src, --dst, regnode);
14247 #ifdef RE_TRACK_PATTERN_OFFSETS
14248 if (RExC_offsets) { /* MJD 20010112 */
14249 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
14253 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
14254 ? "Overwriting end of array!\n" : "OK",
14255 (UV)(src - RExC_emit_start),
14256 (UV)(dst - RExC_emit_start),
14257 (UV)RExC_offsets[0]));
14258 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
14259 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
14265 place = opnd; /* Op node, where operand used to be. */
14266 #ifdef RE_TRACK_PATTERN_OFFSETS
14267 if (RExC_offsets) { /* MJD */
14268 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14272 (UV)(place - RExC_emit_start) > RExC_offsets[0]
14273 ? "Overwriting end of array!\n" : "OK",
14274 (UV)(place - RExC_emit_start),
14275 (UV)(RExC_parse - RExC_start),
14276 (UV)RExC_offsets[0]));
14277 Set_Node_Offset(place, RExC_parse);
14278 Set_Node_Length(place, 1);
14281 src = NEXTOPER(place);
14282 FILL_ADVANCE_NODE(place, op);
14283 Zero(src, offset, regnode);
14287 - regtail - set the next-pointer at the end of a node chain of p to val.
14288 - SEE ALSO: regtail_study
14290 /* TODO: All three parms should be const */
14292 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14296 GET_RE_DEBUG_FLAGS_DECL;
14298 PERL_ARGS_ASSERT_REGTAIL;
14300 PERL_UNUSED_ARG(depth);
14306 /* Find last node. */
14309 regnode * const temp = regnext(scan);
14311 SV * const mysv=sv_newmortal();
14312 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
14313 regprop(RExC_rx, mysv, scan);
14314 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
14315 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
14316 (temp == NULL ? "->" : ""),
14317 (temp == NULL ? PL_reg_name[OP(val)] : "")
14325 if (reg_off_by_arg[OP(scan)]) {
14326 ARG_SET(scan, val - scan);
14329 NEXT_OFF(scan) = val - scan;
14335 - regtail_study - set the next-pointer at the end of a node chain of p to val.
14336 - Look for optimizable sequences at the same time.
14337 - currently only looks for EXACT chains.
14339 This is experimental code. The idea is to use this routine to perform
14340 in place optimizations on branches and groups as they are constructed,
14341 with the long term intention of removing optimization from study_chunk so
14342 that it is purely analytical.
14344 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
14345 to control which is which.
14348 /* TODO: All four parms should be const */
14351 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14356 #ifdef EXPERIMENTAL_INPLACESCAN
14359 GET_RE_DEBUG_FLAGS_DECL;
14361 PERL_ARGS_ASSERT_REGTAIL_STUDY;
14367 /* Find last node. */
14371 regnode * const temp = regnext(scan);
14372 #ifdef EXPERIMENTAL_INPLACESCAN
14373 if (PL_regkind[OP(scan)] == EXACT) {
14374 bool has_exactf_sharp_s; /* Unexamined in this routine */
14375 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
14380 switch (OP(scan)) {
14386 case EXACTFU_TRICKYFOLD:
14388 if( exact == PSEUDO )
14390 else if ( exact != OP(scan) )
14399 SV * const mysv=sv_newmortal();
14400 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
14401 regprop(RExC_rx, mysv, scan);
14402 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
14403 SvPV_nolen_const(mysv),
14404 REG_NODE_NUM(scan),
14405 PL_reg_name[exact]);
14412 SV * const mysv_val=sv_newmortal();
14413 DEBUG_PARSE_MSG("");
14414 regprop(RExC_rx, mysv_val, val);
14415 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
14416 SvPV_nolen_const(mysv_val),
14417 (IV)REG_NODE_NUM(val),
14421 if (reg_off_by_arg[OP(scan)]) {
14422 ARG_SET(scan, val - scan);
14425 NEXT_OFF(scan) = val - scan;
14433 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
14438 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
14443 for (bit=0; bit<32; bit++) {
14444 if (flags & (1<<bit)) {
14445 if (!set++ && lead)
14446 PerlIO_printf(Perl_debug_log, "%s",lead);
14447 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
14452 PerlIO_printf(Perl_debug_log, "\n");
14454 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14459 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
14465 for (bit=0; bit<32; bit++) {
14466 if (flags & (1<<bit)) {
14467 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
14470 if (!set++ && lead)
14471 PerlIO_printf(Perl_debug_log, "%s",lead);
14472 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
14475 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
14476 if (!set++ && lead) {
14477 PerlIO_printf(Perl_debug_log, "%s",lead);
14480 case REGEX_UNICODE_CHARSET:
14481 PerlIO_printf(Perl_debug_log, "UNICODE");
14483 case REGEX_LOCALE_CHARSET:
14484 PerlIO_printf(Perl_debug_log, "LOCALE");
14486 case REGEX_ASCII_RESTRICTED_CHARSET:
14487 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
14489 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
14490 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
14493 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
14499 PerlIO_printf(Perl_debug_log, "\n");
14501 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14507 Perl_regdump(pTHX_ const regexp *r)
14511 SV * const sv = sv_newmortal();
14512 SV *dsv= sv_newmortal();
14513 RXi_GET_DECL(r,ri);
14514 GET_RE_DEBUG_FLAGS_DECL;
14516 PERL_ARGS_ASSERT_REGDUMP;
14518 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
14520 /* Header fields of interest. */
14521 if (r->anchored_substr) {
14522 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
14523 RE_SV_DUMPLEN(r->anchored_substr), 30);
14524 PerlIO_printf(Perl_debug_log,
14525 "anchored %s%s at %"IVdf" ",
14526 s, RE_SV_TAIL(r->anchored_substr),
14527 (IV)r->anchored_offset);
14528 } else if (r->anchored_utf8) {
14529 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
14530 RE_SV_DUMPLEN(r->anchored_utf8), 30);
14531 PerlIO_printf(Perl_debug_log,
14532 "anchored utf8 %s%s at %"IVdf" ",
14533 s, RE_SV_TAIL(r->anchored_utf8),
14534 (IV)r->anchored_offset);
14536 if (r->float_substr) {
14537 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
14538 RE_SV_DUMPLEN(r->float_substr), 30);
14539 PerlIO_printf(Perl_debug_log,
14540 "floating %s%s at %"IVdf"..%"UVuf" ",
14541 s, RE_SV_TAIL(r->float_substr),
14542 (IV)r->float_min_offset, (UV)r->float_max_offset);
14543 } else if (r->float_utf8) {
14544 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
14545 RE_SV_DUMPLEN(r->float_utf8), 30);
14546 PerlIO_printf(Perl_debug_log,
14547 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
14548 s, RE_SV_TAIL(r->float_utf8),
14549 (IV)r->float_min_offset, (UV)r->float_max_offset);
14551 if (r->check_substr || r->check_utf8)
14552 PerlIO_printf(Perl_debug_log,
14554 (r->check_substr == r->float_substr
14555 && r->check_utf8 == r->float_utf8
14556 ? "(checking floating" : "(checking anchored"));
14557 if (r->extflags & RXf_NOSCAN)
14558 PerlIO_printf(Perl_debug_log, " noscan");
14559 if (r->extflags & RXf_CHECK_ALL)
14560 PerlIO_printf(Perl_debug_log, " isall");
14561 if (r->check_substr || r->check_utf8)
14562 PerlIO_printf(Perl_debug_log, ") ");
14564 if (ri->regstclass) {
14565 regprop(r, sv, ri->regstclass);
14566 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
14568 if (r->extflags & RXf_ANCH) {
14569 PerlIO_printf(Perl_debug_log, "anchored");
14570 if (r->extflags & RXf_ANCH_BOL)
14571 PerlIO_printf(Perl_debug_log, "(BOL)");
14572 if (r->extflags & RXf_ANCH_MBOL)
14573 PerlIO_printf(Perl_debug_log, "(MBOL)");
14574 if (r->extflags & RXf_ANCH_SBOL)
14575 PerlIO_printf(Perl_debug_log, "(SBOL)");
14576 if (r->extflags & RXf_ANCH_GPOS)
14577 PerlIO_printf(Perl_debug_log, "(GPOS)");
14578 PerlIO_putc(Perl_debug_log, ' ');
14580 if (r->extflags & RXf_GPOS_SEEN)
14581 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
14582 if (r->intflags & PREGf_SKIP)
14583 PerlIO_printf(Perl_debug_log, "plus ");
14584 if (r->intflags & PREGf_IMPLICIT)
14585 PerlIO_printf(Perl_debug_log, "implicit ");
14586 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
14587 if (r->extflags & RXf_EVAL_SEEN)
14588 PerlIO_printf(Perl_debug_log, "with eval ");
14589 PerlIO_printf(Perl_debug_log, "\n");
14591 regdump_extflags("r->extflags: ",r->extflags);
14592 regdump_intflags("r->intflags: ",r->intflags);
14595 PERL_ARGS_ASSERT_REGDUMP;
14596 PERL_UNUSED_CONTEXT;
14597 PERL_UNUSED_ARG(r);
14598 #endif /* DEBUGGING */
14602 - regprop - printable representation of opcode
14604 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
14607 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
14608 if (flags & ANYOF_INVERT) \
14609 /*make sure the invert info is in each */ \
14610 sv_catpvs(sv, "^"); \
14616 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
14622 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
14623 static const char * const anyofs[] = {
14624 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
14625 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
14626 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
14627 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
14628 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
14629 || _CC_VERTSPACE != 16
14630 #error Need to adjust order of anyofs[]
14667 RXi_GET_DECL(prog,progi);
14668 GET_RE_DEBUG_FLAGS_DECL;
14670 PERL_ARGS_ASSERT_REGPROP;
14674 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
14675 /* It would be nice to FAIL() here, but this may be called from
14676 regexec.c, and it would be hard to supply pRExC_state. */
14677 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
14678 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
14680 k = PL_regkind[OP(o)];
14683 sv_catpvs(sv, " ");
14684 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
14685 * is a crude hack but it may be the best for now since
14686 * we have no flag "this EXACTish node was UTF-8"
14688 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
14689 PERL_PV_ESCAPE_UNI_DETECT |
14690 PERL_PV_ESCAPE_NONASCII |
14691 PERL_PV_PRETTY_ELLIPSES |
14692 PERL_PV_PRETTY_LTGT |
14693 PERL_PV_PRETTY_NOCLEAR
14695 } else if (k == TRIE) {
14696 /* print the details of the trie in dumpuntil instead, as
14697 * progi->data isn't available here */
14698 const char op = OP(o);
14699 const U32 n = ARG(o);
14700 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
14701 (reg_ac_data *)progi->data->data[n] :
14703 const reg_trie_data * const trie
14704 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
14706 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
14707 DEBUG_TRIE_COMPILE_r(
14708 Perl_sv_catpvf(aTHX_ sv,
14709 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
14710 (UV)trie->startstate,
14711 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
14712 (UV)trie->wordcount,
14715 (UV)TRIE_CHARCOUNT(trie),
14716 (UV)trie->uniquecharcount
14719 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
14720 sv_catpvs(sv, "[");
14721 (void) put_latin1_charclass_innards(sv, IS_ANYOF_TRIE(op)
14723 : TRIE_BITMAP(trie));
14724 sv_catpvs(sv, "]");
14727 } else if (k == CURLY) {
14728 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
14729 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
14730 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
14732 else if (k == WHILEM && o->flags) /* Ordinal/of */
14733 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
14734 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
14735 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
14736 if ( RXp_PAREN_NAMES(prog) ) {
14737 if ( k != REF || (OP(o) < NREF)) {
14738 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
14739 SV **name= av_fetch(list, ARG(o), 0 );
14741 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14744 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
14745 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
14746 I32 *nums=(I32*)SvPVX(sv_dat);
14747 SV **name= av_fetch(list, nums[0], 0 );
14750 for ( n=0; n<SvIVX(sv_dat); n++ ) {
14751 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
14752 (n ? "," : ""), (IV)nums[n]);
14754 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14758 } else if (k == GOSUB)
14759 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
14760 else if (k == VERB) {
14762 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
14763 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
14764 } else if (k == LOGICAL)
14765 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
14766 else if (k == ANYOF) {
14767 const U8 flags = ANYOF_FLAGS(o);
14771 if (flags & ANYOF_LOCALE)
14772 sv_catpvs(sv, "{loc}");
14773 if (flags & ANYOF_LOC_FOLD)
14774 sv_catpvs(sv, "{i}");
14775 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
14776 if (flags & ANYOF_INVERT)
14777 sv_catpvs(sv, "^");
14779 /* output what the standard cp 0-255 bitmap matches */
14780 do_sep = put_latin1_charclass_innards(sv, ANYOF_BITMAP(o));
14782 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14783 /* output any special charclass tests (used entirely under use locale) */
14784 if (ANYOF_CLASS_TEST_ANY_SET(o)) {
14786 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++) {
14787 if (ANYOF_CLASS_TEST(o,i)) {
14788 sv_catpv(sv, anyofs[i]);
14794 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14796 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
14797 sv_catpvs(sv, "{non-utf8-latin1-all}");
14800 /* output information about the unicode matching */
14801 if (flags & ANYOF_UNICODE_ALL)
14802 sv_catpvs(sv, "{unicode_all}");
14803 else if (ANYOF_NONBITMAP(o)) {
14804 SV *lv; /* Set if there is something outside the bit map. */
14806 bool byte_output = FALSE; /* If something in the bitmap has been
14809 if (flags & ANYOF_NONBITMAP_NON_UTF8) {
14810 sv_catpvs(sv, "{outside bitmap}");
14813 sv_catpvs(sv, "{utf8}");
14816 /* Get the stuff that wasn't in the bitmap */
14817 sw = regclass_swash(prog, o, FALSE, &lv, NULL);
14818 if (lv && lv != &PL_sv_undef) {
14819 char *s = savesvpv(lv);
14820 char * const origs = s;
14822 while (*s && *s != '\n')
14826 const char * const t = ++s;
14829 sv_catpvs(sv, " ");
14835 /* Truncate very long output */
14836 if (s - origs > 256) {
14837 Perl_sv_catpvf(aTHX_ sv,
14839 (int) (s - origs - 1),
14845 else if (*s == '\t') {
14859 SvREFCNT_dec_NN(lv);
14863 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
14865 else if (k == POSIXD || k == NPOSIXD) {
14866 U8 index = FLAGS(o) * 2;
14867 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
14868 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
14871 sv_catpv(sv, anyofs[index]);
14874 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
14875 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
14877 PERL_UNUSED_CONTEXT;
14878 PERL_UNUSED_ARG(sv);
14879 PERL_UNUSED_ARG(o);
14880 PERL_UNUSED_ARG(prog);
14881 #endif /* DEBUGGING */
14885 Perl_re_intuit_string(pTHX_ REGEXP * const r)
14886 { /* Assume that RE_INTUIT is set */
14888 struct regexp *const prog = ReANY(r);
14889 GET_RE_DEBUG_FLAGS_DECL;
14891 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
14892 PERL_UNUSED_CONTEXT;
14896 const char * const s = SvPV_nolen_const(prog->check_substr
14897 ? prog->check_substr : prog->check_utf8);
14899 if (!PL_colorset) reginitcolors();
14900 PerlIO_printf(Perl_debug_log,
14901 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
14903 prog->check_substr ? "" : "utf8 ",
14904 PL_colors[5],PL_colors[0],
14907 (strlen(s) > 60 ? "..." : ""));
14910 return prog->check_substr ? prog->check_substr : prog->check_utf8;
14916 handles refcounting and freeing the perl core regexp structure. When
14917 it is necessary to actually free the structure the first thing it
14918 does is call the 'free' method of the regexp_engine associated to
14919 the regexp, allowing the handling of the void *pprivate; member
14920 first. (This routine is not overridable by extensions, which is why
14921 the extensions free is called first.)
14923 See regdupe and regdupe_internal if you change anything here.
14925 #ifndef PERL_IN_XSUB_RE
14927 Perl_pregfree(pTHX_ REGEXP *r)
14933 Perl_pregfree2(pTHX_ REGEXP *rx)
14936 struct regexp *const r = ReANY(rx);
14937 GET_RE_DEBUG_FLAGS_DECL;
14939 PERL_ARGS_ASSERT_PREGFREE2;
14941 if (r->mother_re) {
14942 ReREFCNT_dec(r->mother_re);
14944 CALLREGFREE_PVT(rx); /* free the private data */
14945 SvREFCNT_dec(RXp_PAREN_NAMES(r));
14946 Safefree(r->xpv_len_u.xpvlenu_pv);
14949 SvREFCNT_dec(r->anchored_substr);
14950 SvREFCNT_dec(r->anchored_utf8);
14951 SvREFCNT_dec(r->float_substr);
14952 SvREFCNT_dec(r->float_utf8);
14953 Safefree(r->substrs);
14955 RX_MATCH_COPY_FREE(rx);
14956 #ifdef PERL_ANY_COW
14957 SvREFCNT_dec(r->saved_copy);
14960 SvREFCNT_dec(r->qr_anoncv);
14961 rx->sv_u.svu_rx = 0;
14966 This is a hacky workaround to the structural issue of match results
14967 being stored in the regexp structure which is in turn stored in
14968 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
14969 could be PL_curpm in multiple contexts, and could require multiple
14970 result sets being associated with the pattern simultaneously, such
14971 as when doing a recursive match with (??{$qr})
14973 The solution is to make a lightweight copy of the regexp structure
14974 when a qr// is returned from the code executed by (??{$qr}) this
14975 lightweight copy doesn't actually own any of its data except for
14976 the starp/end and the actual regexp structure itself.
14982 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
14984 struct regexp *ret;
14985 struct regexp *const r = ReANY(rx);
14986 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
14988 PERL_ARGS_ASSERT_REG_TEMP_COPY;
14991 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
14993 SvOK_off((SV *)ret_x);
14995 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
14996 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
14997 made both spots point to the same regexp body.) */
14998 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
14999 assert(!SvPVX(ret_x));
15000 ret_x->sv_u.svu_rx = temp->sv_any;
15001 temp->sv_any = NULL;
15002 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
15003 SvREFCNT_dec_NN(temp);
15004 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
15005 ing below will not set it. */
15006 SvCUR_set(ret_x, SvCUR(rx));
15009 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
15010 sv_force_normal(sv) is called. */
15012 ret = ReANY(ret_x);
15014 SvFLAGS(ret_x) |= SvUTF8(rx);
15015 /* We share the same string buffer as the original regexp, on which we
15016 hold a reference count, incremented when mother_re is set below.
15017 The string pointer is copied here, being part of the regexp struct.
15019 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
15020 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
15022 const I32 npar = r->nparens+1;
15023 Newx(ret->offs, npar, regexp_paren_pair);
15024 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15027 Newx(ret->substrs, 1, struct reg_substr_data);
15028 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15030 SvREFCNT_inc_void(ret->anchored_substr);
15031 SvREFCNT_inc_void(ret->anchored_utf8);
15032 SvREFCNT_inc_void(ret->float_substr);
15033 SvREFCNT_inc_void(ret->float_utf8);
15035 /* check_substr and check_utf8, if non-NULL, point to either their
15036 anchored or float namesakes, and don't hold a second reference. */
15038 RX_MATCH_COPIED_off(ret_x);
15039 #ifdef PERL_ANY_COW
15040 ret->saved_copy = NULL;
15042 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
15043 SvREFCNT_inc_void(ret->qr_anoncv);
15049 /* regfree_internal()
15051 Free the private data in a regexp. This is overloadable by
15052 extensions. Perl takes care of the regexp structure in pregfree(),
15053 this covers the *pprivate pointer which technically perl doesn't
15054 know about, however of course we have to handle the
15055 regexp_internal structure when no extension is in use.
15057 Note this is called before freeing anything in the regexp
15062 Perl_regfree_internal(pTHX_ REGEXP * const rx)
15065 struct regexp *const r = ReANY(rx);
15066 RXi_GET_DECL(r,ri);
15067 GET_RE_DEBUG_FLAGS_DECL;
15069 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
15075 SV *dsv= sv_newmortal();
15076 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
15077 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
15078 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
15079 PL_colors[4],PL_colors[5],s);
15082 #ifdef RE_TRACK_PATTERN_OFFSETS
15084 Safefree(ri->u.offsets); /* 20010421 MJD */
15086 if (ri->code_blocks) {
15088 for (n = 0; n < ri->num_code_blocks; n++)
15089 SvREFCNT_dec(ri->code_blocks[n].src_regex);
15090 Safefree(ri->code_blocks);
15094 int n = ri->data->count;
15097 /* If you add a ->what type here, update the comment in regcomp.h */
15098 switch (ri->data->what[n]) {
15104 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
15107 Safefree(ri->data->data[n]);
15113 { /* Aho Corasick add-on structure for a trie node.
15114 Used in stclass optimization only */
15116 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
15118 refcount = --aho->refcount;
15121 PerlMemShared_free(aho->states);
15122 PerlMemShared_free(aho->fail);
15123 /* do this last!!!! */
15124 PerlMemShared_free(ri->data->data[n]);
15125 PerlMemShared_free(ri->regstclass);
15131 /* trie structure. */
15133 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
15135 refcount = --trie->refcount;
15138 PerlMemShared_free(trie->charmap);
15139 PerlMemShared_free(trie->states);
15140 PerlMemShared_free(trie->trans);
15142 PerlMemShared_free(trie->bitmap);
15144 PerlMemShared_free(trie->jump);
15145 PerlMemShared_free(trie->wordinfo);
15146 /* do this last!!!! */
15147 PerlMemShared_free(ri->data->data[n]);
15152 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
15155 Safefree(ri->data->what);
15156 Safefree(ri->data);
15162 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
15163 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
15164 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
15167 re_dup - duplicate a regexp.
15169 This routine is expected to clone a given regexp structure. It is only
15170 compiled under USE_ITHREADS.
15172 After all of the core data stored in struct regexp is duplicated
15173 the regexp_engine.dupe method is used to copy any private data
15174 stored in the *pprivate pointer. This allows extensions to handle
15175 any duplication it needs to do.
15177 See pregfree() and regfree_internal() if you change anything here.
15179 #if defined(USE_ITHREADS)
15180 #ifndef PERL_IN_XSUB_RE
15182 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
15186 const struct regexp *r = ReANY(sstr);
15187 struct regexp *ret = ReANY(dstr);
15189 PERL_ARGS_ASSERT_RE_DUP_GUTS;
15191 npar = r->nparens+1;
15192 Newx(ret->offs, npar, regexp_paren_pair);
15193 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15195 if (ret->substrs) {
15196 /* Do it this way to avoid reading from *r after the StructCopy().
15197 That way, if any of the sv_dup_inc()s dislodge *r from the L1
15198 cache, it doesn't matter. */
15199 const bool anchored = r->check_substr
15200 ? r->check_substr == r->anchored_substr
15201 : r->check_utf8 == r->anchored_utf8;
15202 Newx(ret->substrs, 1, struct reg_substr_data);
15203 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15205 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
15206 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
15207 ret->float_substr = sv_dup_inc(ret->float_substr, param);
15208 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
15210 /* check_substr and check_utf8, if non-NULL, point to either their
15211 anchored or float namesakes, and don't hold a second reference. */
15213 if (ret->check_substr) {
15215 assert(r->check_utf8 == r->anchored_utf8);
15216 ret->check_substr = ret->anchored_substr;
15217 ret->check_utf8 = ret->anchored_utf8;
15219 assert(r->check_substr == r->float_substr);
15220 assert(r->check_utf8 == r->float_utf8);
15221 ret->check_substr = ret->float_substr;
15222 ret->check_utf8 = ret->float_utf8;
15224 } else if (ret->check_utf8) {
15226 ret->check_utf8 = ret->anchored_utf8;
15228 ret->check_utf8 = ret->float_utf8;
15233 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
15234 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
15237 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
15239 if (RX_MATCH_COPIED(dstr))
15240 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
15242 ret->subbeg = NULL;
15243 #ifdef PERL_ANY_COW
15244 ret->saved_copy = NULL;
15247 /* Whether mother_re be set or no, we need to copy the string. We
15248 cannot refrain from copying it when the storage points directly to
15249 our mother regexp, because that's
15250 1: a buffer in a different thread
15251 2: something we no longer hold a reference on
15252 so we need to copy it locally. */
15253 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
15254 ret->mother_re = NULL;
15256 #endif /* PERL_IN_XSUB_RE */
15261 This is the internal complement to regdupe() which is used to copy
15262 the structure pointed to by the *pprivate pointer in the regexp.
15263 This is the core version of the extension overridable cloning hook.
15264 The regexp structure being duplicated will be copied by perl prior
15265 to this and will be provided as the regexp *r argument, however
15266 with the /old/ structures pprivate pointer value. Thus this routine
15267 may override any copying normally done by perl.
15269 It returns a pointer to the new regexp_internal structure.
15273 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
15276 struct regexp *const r = ReANY(rx);
15277 regexp_internal *reti;
15279 RXi_GET_DECL(r,ri);
15281 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
15285 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
15286 Copy(ri->program, reti->program, len+1, regnode);
15288 reti->num_code_blocks = ri->num_code_blocks;
15289 if (ri->code_blocks) {
15291 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
15292 struct reg_code_block);
15293 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
15294 struct reg_code_block);
15295 for (n = 0; n < ri->num_code_blocks; n++)
15296 reti->code_blocks[n].src_regex = (REGEXP*)
15297 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
15300 reti->code_blocks = NULL;
15302 reti->regstclass = NULL;
15305 struct reg_data *d;
15306 const int count = ri->data->count;
15309 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
15310 char, struct reg_data);
15311 Newx(d->what, count, U8);
15314 for (i = 0; i < count; i++) {
15315 d->what[i] = ri->data->what[i];
15316 switch (d->what[i]) {
15317 /* see also regcomp.h and regfree_internal() */
15318 case 'a': /* actually an AV, but the dup function is identical. */
15322 case 'u': /* actually an HV, but the dup function is identical. */
15323 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
15326 /* This is cheating. */
15327 Newx(d->data[i], 1, struct regnode_charclass_class);
15328 StructCopy(ri->data->data[i], d->data[i],
15329 struct regnode_charclass_class);
15330 reti->regstclass = (regnode*)d->data[i];
15333 /* Trie stclasses are readonly and can thus be shared
15334 * without duplication. We free the stclass in pregfree
15335 * when the corresponding reg_ac_data struct is freed.
15337 reti->regstclass= ri->regstclass;
15341 ((reg_trie_data*)ri->data->data[i])->refcount++;
15346 d->data[i] = ri->data->data[i];
15349 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
15358 reti->name_list_idx = ri->name_list_idx;
15360 #ifdef RE_TRACK_PATTERN_OFFSETS
15361 if (ri->u.offsets) {
15362 Newx(reti->u.offsets, 2*len+1, U32);
15363 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
15366 SetProgLen(reti,len);
15369 return (void*)reti;
15372 #endif /* USE_ITHREADS */
15374 #ifndef PERL_IN_XSUB_RE
15377 - regnext - dig the "next" pointer out of a node
15380 Perl_regnext(pTHX_ regnode *p)
15388 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
15389 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
15392 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
15401 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
15404 STRLEN l1 = strlen(pat1);
15405 STRLEN l2 = strlen(pat2);
15408 const char *message;
15410 PERL_ARGS_ASSERT_RE_CROAK2;
15416 Copy(pat1, buf, l1 , char);
15417 Copy(pat2, buf + l1, l2 , char);
15418 buf[l1 + l2] = '\n';
15419 buf[l1 + l2 + 1] = '\0';
15421 /* ANSI variant takes additional second argument */
15422 va_start(args, pat2);
15426 msv = vmess(buf, &args);
15428 message = SvPV_const(msv,l1);
15431 Copy(message, buf, l1 , char);
15432 buf[l1-1] = '\0'; /* Overwrite \n */
15433 Perl_croak(aTHX_ "%s", buf);
15436 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
15438 #ifndef PERL_IN_XSUB_RE
15440 Perl_save_re_context(pTHX)
15444 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
15446 const REGEXP * const rx = PM_GETRE(PL_curpm);
15449 for (i = 1; i <= RX_NPARENS(rx); i++) {
15450 char digits[TYPE_CHARS(long)];
15451 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
15452 GV *const *const gvp
15453 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
15456 GV * const gv = *gvp;
15457 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
15469 S_put_byte(pTHX_ SV *sv, int c)
15471 PERL_ARGS_ASSERT_PUT_BYTE;
15473 /* Our definition of isPRINT() ignores locales, so only bytes that are
15474 not part of UTF-8 are considered printable. I assume that the same
15475 holds for UTF-EBCDIC.
15476 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
15477 which Wikipedia says:
15479 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
15480 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
15481 identical, to the ASCII delete (DEL) or rubout control character. ...
15482 it is typically mapped to hexadecimal code 9F, in order to provide a
15483 unique character mapping in both directions)
15485 So the old condition can be simplified to !isPRINT(c) */
15488 case '\r': Perl_sv_catpvf(aTHX_ sv, "\\r"); break;
15489 case '\n': Perl_sv_catpvf(aTHX_ sv, "\\n"); break;
15490 case '\t': Perl_sv_catpvf(aTHX_ sv, "\\t"); break;
15491 case '\f': Perl_sv_catpvf(aTHX_ sv, "\\f"); break;
15492 case '\a': Perl_sv_catpvf(aTHX_ sv, "\\a"); break;
15495 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
15500 const char string = c;
15501 if (c == '-' || c == ']' || c == '\\' || c == '^')
15502 sv_catpvs(sv, "\\");
15503 sv_catpvn(sv, &string, 1);
15508 S_put_latin1_charclass_innards(pTHX_ SV *sv, char *bitmap)
15510 /* Appends to 'sv' a displayable version of the innards of the bracketed
15511 * character class whose bitmap is 'bitmap'; Returns 'TRUE' if it actually
15512 * output anything */
15515 int rangestart = -1;
15516 bool has_output_anything = FALSE;
15518 PERL_ARGS_ASSERT_PUT_LATIN1_CHARCLASS_INNARDS;
15520 for (i = 0; i <= 256; i++) {
15521 if (i < 256 && BITMAP_TEST((U8 *) bitmap,i)) {
15522 if (rangestart == -1)
15524 } else if (rangestart != -1) {
15526 if (i <= rangestart + 3) { /* Individual chars in short ranges */
15527 for (; rangestart < i; rangestart++)
15528 put_byte(sv, rangestart);
15531 || ! isALPHANUMERIC(rangestart)
15532 || ! isALPHANUMERIC(j)
15533 || isDIGIT(rangestart) != isDIGIT(j)
15534 || isUPPER(rangestart) != isUPPER(j)
15535 || isLOWER(rangestart) != isLOWER(j)
15537 /* This final test should get optimized out except
15538 * on EBCDIC platforms, where it causes ranges that
15539 * cross discontinuities like i/j to be shown as hex
15540 * instead of the misleading, e.g. H-K (since that
15541 * range includes more than H, I, J, K). */
15542 || (j - rangestart)
15543 != NATIVE_TO_ASCII(j) - NATIVE_TO_ASCII(rangestart))
15545 Perl_sv_catpvf(aTHX_ sv, "\\x{%02x}-\\x{%02x}",
15547 (j < 256) ? j : 255);
15549 else { /* Here, the ends of the range are both digits, or both
15550 uppercase, or both lowercase; and there's no
15551 discontinuity in the range (which could happen on EBCDIC
15553 put_byte(sv, rangestart);
15554 sv_catpvs(sv, "-");
15558 has_output_anything = TRUE;
15562 return has_output_anything;
15565 #define CLEAR_OPTSTART \
15566 if (optstart) STMT_START { \
15567 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
15571 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
15573 STATIC const regnode *
15574 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
15575 const regnode *last, const regnode *plast,
15576 SV* sv, I32 indent, U32 depth)
15579 U8 op = PSEUDO; /* Arbitrary non-END op. */
15580 const regnode *next;
15581 const regnode *optstart= NULL;
15583 RXi_GET_DECL(r,ri);
15584 GET_RE_DEBUG_FLAGS_DECL;
15586 PERL_ARGS_ASSERT_DUMPUNTIL;
15588 #ifdef DEBUG_DUMPUNTIL
15589 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
15590 last ? last-start : 0,plast ? plast-start : 0);
15593 if (plast && plast < last)
15596 while (PL_regkind[op] != END && (!last || node < last)) {
15597 /* While that wasn't END last time... */
15600 if (op == CLOSE || op == WHILEM)
15602 next = regnext((regnode *)node);
15605 if (OP(node) == OPTIMIZED) {
15606 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
15613 regprop(r, sv, node);
15614 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
15615 (int)(2*indent + 1), "", SvPVX_const(sv));
15617 if (OP(node) != OPTIMIZED) {
15618 if (next == NULL) /* Next ptr. */
15619 PerlIO_printf(Perl_debug_log, " (0)");
15620 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
15621 PerlIO_printf(Perl_debug_log, " (FAIL)");
15623 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
15624 (void)PerlIO_putc(Perl_debug_log, '\n');
15628 if (PL_regkind[(U8)op] == BRANCHJ) {
15631 const regnode *nnode = (OP(next) == LONGJMP
15632 ? regnext((regnode *)next)
15634 if (last && nnode > last)
15636 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
15639 else if (PL_regkind[(U8)op] == BRANCH) {
15641 DUMPUNTIL(NEXTOPER(node), next);
15643 else if ( PL_regkind[(U8)op] == TRIE ) {
15644 const regnode *this_trie = node;
15645 const char op = OP(node);
15646 const U32 n = ARG(node);
15647 const reg_ac_data * const ac = op>=AHOCORASICK ?
15648 (reg_ac_data *)ri->data->data[n] :
15650 const reg_trie_data * const trie =
15651 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
15653 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
15655 const regnode *nextbranch= NULL;
15658 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
15659 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
15661 PerlIO_printf(Perl_debug_log, "%*s%s ",
15662 (int)(2*(indent+3)), "",
15663 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
15664 PL_colors[0], PL_colors[1],
15665 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
15666 PERL_PV_PRETTY_ELLIPSES |
15667 PERL_PV_PRETTY_LTGT
15672 U16 dist= trie->jump[word_idx+1];
15673 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
15674 (UV)((dist ? this_trie + dist : next) - start));
15677 nextbranch= this_trie + trie->jump[0];
15678 DUMPUNTIL(this_trie + dist, nextbranch);
15680 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
15681 nextbranch= regnext((regnode *)nextbranch);
15683 PerlIO_printf(Perl_debug_log, "\n");
15686 if (last && next > last)
15691 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
15692 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
15693 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
15695 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
15697 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
15699 else if ( op == PLUS || op == STAR) {
15700 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
15702 else if (PL_regkind[(U8)op] == ANYOF) {
15703 /* arglen 1 + class block */
15704 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
15705 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
15706 node = NEXTOPER(node);
15708 else if (PL_regkind[(U8)op] == EXACT) {
15709 /* Literal string, where present. */
15710 node += NODE_SZ_STR(node) - 1;
15711 node = NEXTOPER(node);
15714 node = NEXTOPER(node);
15715 node += regarglen[(U8)op];
15717 if (op == CURLYX || op == OPEN)
15721 #ifdef DEBUG_DUMPUNTIL
15722 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
15727 #endif /* DEBUGGING */
15731 * c-indentation-style: bsd
15732 * c-basic-offset: 4
15733 * indent-tabs-mode: nil
15736 * ex: set ts=8 sts=4 sw=4 et: